Intelligent wearable device identity authentication method and system based on biological characteristics

Abstract

The invention relates to the technical field of biological recognition, in particular to an intelligent wearable device identity authentication method and system based on biological characteristics, and the method comprises the following steps: collecting a user voice sample, removing background noise from the voice sample through a spectral subtraction method and a Wiener filter technology, and generating a denoised voice sample; separating the time sequence and frequency spectrum information of the voice signal from the de-noised voice sample by using a convolutional neural network, and extracting a time frequency spectrum feature vector; according to the method, background noise is eliminated from a voice sample by using spectral subtraction and a Wiener filter, and time spectrum features are extracted in combination with the convolutional neural network, so that the user voice capturing and analyzing capability is improved. In addition, the gating circulation unit deeply learns the intonation and rhythm of the voice data, so that the recognition of the voice mode is more accurate. The sensor data is subjected to independence analysis through the isolated forest algorithm, abnormal data can be recognized and corrected, and the accuracy of data output is ensured.

Description

Technical Field

[0001] The present invention relates to the field of biometric identification technology, and in particular to a method and system for authenticating an intelligent wearable device based on biometric features. Background Art

[0002] Biometrics is a technology field that uses an individual's physiological or behavioral characteristics for identity verification. These characteristics include, but are not limited to, fingerprints, faces, irises, voices, gestures, and gait. Because these characteristics are unique and difficult to copy, biometrics provides a highly secure and convenient way to authenticate users. However, existing smart wearable devices rely on fingerprint or facial recognition, which may cause verification failures under certain conditions, such as insufficient light. In addition, it is difficult to dynamically adjust abnormal data, which limits processing capabilities when data quality is compromised. Therefore, improvements are needed. Summary of the invention

[0003] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a biometric-based smart wearable device identity authentication method and system.

[0004] In order to achieve the above object, the present invention adopts the following technical solution, a method for identity authentication of a smart wearable device based on biometrics, comprising the following steps: Collect user voice samples, remove background noise from the voice samples using spectral subtraction and Wiener filter technology to generate denoised voice samples; use convolutional neural network to separate the time series and spectrum information of the voice signal from the denoised voice samples, and extract the time spectrum feature vector; Input the time-frequency spectrum feature vector into the gated recurrent unit, strengthen the learning of intonation and rhythm in the speech data through sequence analysis, and generate a speech dynamic analysis model; verify the user identity through the speech dynamic analysis model to obtain a user identity verification result; Based on the user identity authentication result, the sensor data stream of the smart wearable device is collected, and the isolation forest algorithm is used to perform independence analysis on each data point in the data stream, identify the data points that deviate from the normal range, and generate abnormal data identification results; based on the abnormal data identification results, the data output is adjusted according to the normal data, the abnormal data is corrected, and the automatic correction data result is output; Adaptive filtering is applied to the automatic correction data results, and the automatic correction data results are verified using cross-validation to obtain equipment health status results.

[0005] Preferably, the steps of obtaining the denoised speech sample are: Collect user speech samples, apply frequency domain analysis through spectral subtraction, identify interference from non-speech background noise, and generate preliminary denoised speech samples; Based on the preliminary denoised speech sample, a Wiener filter is used for processing, and the speech signal is optimized by spectrum reconstruction to obtain a further denoised speech sample; Based on the further denoised speech sample, sound quality detection and adjustment are performed, and signal quality evaluation is performed to obtain a denoised speech sample.

[0006] Preferably, the step of acquiring the time-frequency spectrum feature vector is: Performing deep feature analysis on the denoised speech sample, processing the speech signal using a convolutional neural network, separating the time series and spectrum information in the speech signal through a convolution layer and a pooling layer, and obtaining time series and spectrum separation data; Based on the time series and spectrum separation data, a multi-layer perceptron is used to analyze and identify the relationship between each time point and the frequency component, extract key time spectrum features, and generate a time spectrum feature vector.

[0007] Preferably, the steps of acquiring the speech dynamic analysis model are: Input the time-frequency spectrum feature vector into a gated recurrent unit, perform initial sequence analysis on the input data, convert it into a serialized format, and generate serialized feature data; Based on the serialized feature data, the intonation and rhythm in the speech data are learned and analyzed, and the feature processing process is repeatedly iterated through loop connections to adjust the data representation and build a speech dynamic analysis model.

[0008] Preferably, the steps of obtaining the user identity verification result are: Using the speech dynamic analysis model to process the time spectrum feature vector, identify the user's speech pattern, and generate preliminary identity authentication data; Based on the preliminary identity verification data, statistical analysis is performed to calculate the identity verification rate, and the calculation formula is: ; in, represents the identification rate, is the user feature vector elements, is the first feature vector The average value of the elements, For the The standard deviation of the elements, represents the contribution rate of each element to identity recognition, and m represents the total number of elements in the feature vector; Based on the identity recognition rate, the degree of matching with the preset standard is evaluated. If the identity recognition rate exceeds the preset standard, the user identity is confirmed, otherwise access is denied, and the user identity verification result is obtained.

[0009] Preferably, the steps for obtaining the abnormal data identification result are: Based on the user identity authentication result, activate the sensor of the smart wearable device, collect physiological and environmental data, and generate a real-time sensor data stream; Based on the real-time sensor data stream, the isolation forest algorithm is applied to perform anomaly detection and calculate the anomaly index of each data point. The calculation formula is: ; in, Represents data points The abnormal indicators It is a data point In the The eigenvalue of dimension, is the importance coefficient of the dimension feature, is the characteristic dimension; Based on the abnormal index, the degree of deviation of each data point from the normal range is evaluated, and the data points with abnormal indexes higher than the threshold are identified as abnormal, and an abnormal data identification result is generated.

[0010] Preferably, the steps of obtaining the automatic correction data result are: Based on the abnormal data identification result, a Kalman filter algorithm is applied to identify abnormal data points, and the abnormal data points are separated from the data stream of the smart wearable device to generate a data stream that has been preliminarily cleaned; Based on the preliminarily cleaned data stream, the smoothing value of each data point is calculated using the following formula: ; in, Representative data points The smoothed value of and Represents data points Data points before and after, and is the weight factor, Represents the range of data points; Based on the smoothed value, the data stream is readjusted to obtain an automatically corrected data result.

[0011] Preferably, the steps of obtaining the equipment health status result are: Applying adaptive filtering to process the automatic correction data results to obtain an optimized data stream; Based on the optimized data stream, by comparing the performance between the data segmentation sets, the operating status of the smart wearable device is analyzed and evaluated, the health status of the device is determined, and the device health status result is generated.

[0012] The present invention provides an identity authentication system, comprising: The voiceprint collection module receives voice samples from users, extracts time series and spectrum information from them, and generates time-spectrum feature vectors; The voiceprint analysis module inputs the time spectrum feature vector into the sequence processing unit, analyzes the intonation and rhythm, and generates a speech dynamic analysis model; The identity verification module uses the voice dynamic analysis model to compare the stored user model, verify the voice consistency, and obtain the user identity verification result; The data anomaly detection module analyzes the sensor data stream of the smart wearable device based on the user identity authentication result, identifies abnormal data points, and generates abnormal data identification results; The data correction and health assessment module adjusts and corrects the data output based on the abnormal data identification results, applies the verification process to the automatically corrected data, and obtains the equipment health status results.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are: The present invention uses spectral subtraction and Wiener filter to remove background noise from speech samples, and combines convolutional neural network to extract time spectrum features, thereby improving the ability to capture and analyze user voice. In addition, the gated recurrent unit conducts in-depth learning of the intonation and rhythm of speech data, making the recognition of speech patterns more accurate. The isolation forest algorithm performs independence analysis on sensor data, can identify and correct abnormal data, and ensure the accuracy of data output. Through adaptive filtering and cross-validation technology, the accuracy of data processing is improved and the health status assessment of the equipment is optimized. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the steps of the present invention; Figure 2 A schematic diagram of the steps for obtaining denoised speech samples in the present invention; Figure 3 Schematic diagram of the steps for obtaining the time-frequency spectrum feature vector in the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0016] See also Figure 1-3 The present invention provides a technical solution, a method for authenticating an intelligent wearable device based on biometrics, comprising the following steps: Collect user voice samples, remove background noise from voice samples using spectral subtraction and Wiener filter technology to generate denoised voice samples; use convolutional neural network to separate the time series and spectrum information of voice signals from denoised voice samples, and extract time-spectrum feature vectors; Input the time spectrum feature vector to the gated recurrent unit, strengthen the learning of intonation and rhythm in the speech data through sequence analysis, and generate a speech dynamic analysis model; verify the user identity through the speech dynamic analysis model to obtain the user identity verification result; Based on the user identity authentication results, the sensor data stream of the smart wearable device is collected, and the isolation forest algorithm is used to perform independence analysis on each data point in the data stream, identify the data points that deviate from the normal range, and generate abnormal data identification results; based on the abnormal data identification results, the data output is adjusted according to the normal data, the abnormal data is corrected, and the automatic correction data results are output; Adaptive filtering is applied to the auto-corrected data results, and cross-validation is used to validate the auto-corrected data results to obtain the equipment health status results.

[0017] The steps for obtaining denoised speech samples are as follows: Collect user speech samples, apply frequency domain analysis through spectral subtraction, identify interference from non-speech background noise, and generate preliminary denoised speech samples; Based on the preliminary denoised speech samples, the Wiener filter is used to process and the speech signal is optimized by spectrum reconstruction to obtain further denoised speech samples; Based on the further denoised speech sample, sound quality detection and adjustment are performed, and signal quality evaluation is performed to obtain a denoised speech sample.

[0018] Specifically, first refer to the audio sampling and background noise benchmark data, and set a noise energy threshold according to the average noise energy level obtained by continuous recording in the actual sampling environment. This threshold is derived from the spectrum statistics of the noise signals collected in different scenes and the average value thereof. Then, the threshold is used as the basic parameter for determining the non-speech components in the calculation process. Then, the collected original speech signal is segmented by using a fixed frame length slicing method. For example, 1024 samples are taken per frame and an overlapping area of ​​512 samples is set between adjacent frames. The segmented signal is subjected to a fast Fourier transform to obtain the frequency components of each frequency. The estimated noise energy is subtracted from the energy of these frequency bands by spectral subtraction to achieve the process of weakening the noise component. At the same time, the relationship between the residual energy of each frequency band after noise reduction and the threshold is compared. If the residual energy is lower than the threshold, it is determined as invalid non-speech interference and the component is removed during inverse transformation, thereby reducing the proportion of unnecessary background noise. The abnormal frames in the analysis process are further checked according to the preset effective range. For example, the frames whose average energy deviates from the unit interval of 0 to 100 are marked as abnormal. Then all the processed segmented data are summarized to reconstruct the time domain signal to obtain a preliminary denoised speech sample.

[0019] After obtaining the preliminary denoised speech sample, the gain coefficient of the Wiener filter is calculated based on the previous statistical results of the noise power spectrum and the speech power spectrum. ,in It can be obtained by comparing the energy distribution of speech and noise at different frequencies. The specific method is to first record the energy mean of speech in each frequency band and the energy mean of noise in the same frequency band over a period of time, set the ratio of these two energy means as the reference of filter gain, and obtain a filter step length based on experience or experimental measurement. ,when If it is larger, it means that the corresponding frequency band is dominated by speech components. If it is too small, it means that the frequency band is dominated by noise. It is adjusted continuously according to the energy ratio obtained previously. And observe whether the gain oscillation occurs. If it is still in a state of frequent fluctuation after repeated iterations, reduce it appropriately. A smoother adjustment is made, and after the parameters are stabilized, the filtered speech signal is obtained through spectrum reconstruction. During this period, if it is found that the energy of some frequency bands still exceeds the previously measured noise mean plus a 10% redundant interval, it is regarded as a potential noise residue and the gain coefficient of the corresponding frequency band is adjusted again. Finally, after multiple iterations, the ideal spectrum balance is achieved to obtain a further denoised speech sample.

[0020] After obtaining the further denoised speech sample, the distortion degree contained in each frequency band is first detected by referring to the sound quality measurement rules retained previously. The sound quality distortion degree is expressed by dividing it into several levels in the range of 0 to 1. For example, less than 0.2 is regarded as the extremely low distortion range, 0.2 to 0.5 is regarded as the medium distortion range, and greater than 0.5 is judged as the high distortion range. If the distortion index corresponding to some time periods exceeds 0.5, it is fine-tuned in combination with the waveform change trend of the current speech signal in the time domain. In this adjustment process, linear interpolation can be used to interpolate and trim the frames with mutations. Then, the signal quality is evaluated according to the average energy distribution, distortion metric and inter-frame smoothing degree obtained previously. During the evaluation, the distortion index of each segmented speech is compared with the preset quality judgment threshold. If most of the distortion indexes are less than 0.3, it is determined that the speech sample has reached the quality standard. If there are still many frames falling into the range of 0.3 to 0.5, the distortion index is further adjusted and recalculated for these frames. After completing all the evaluations, the trimming and detection results are integrated together to obtain the denoised speech sample.

[0021] The steps to obtain the time spectrum feature vector are: Perform deep feature analysis on denoised speech samples, use convolutional neural networks to process speech signals, separate the time series and spectrum information in speech signals through convolutional layers and pooling layers, and obtain time series and spectrum separation data; Based on the time series and spectrum separation data, a multi-layer perceptron is used to analyze and identify the relationship between each time point and the frequency component, extract the key time-spectrum features, and generate the time-spectrum feature vector.

[0022] Specifically, the denoised speech samples are subjected to deep feature analysis. First, the batch size is set to 32 according to the previously recorded speech data scale and sampling frame length, and the learning rate is selected to be 0.001. A training set containing several labeled samples is prepared and the validation set and test set are distinguished. Then, the speech data in the training set is read and organized into a multi-channel input format. When the convolutional neural network is used to process the speech signal, several convolution kernel sizes are divided to cover different time windows and frequency ranges. When defining the number of convolution kernels, the previously accumulated sample diversity statistics are referred to. At the beginning, 64 convolution kernels can be set as the initial value. If the convergence process of the model is observed to be slow after several training rounds, the number of convolution kernels is appropriately increased and the training error changes are recorded. Each convolution operation will output the corresponding feature map. When downsampling the feature map through the pooling layer, 2×2 can be selected. The pooling window is set to 1×2 and the stride is set to 2. If the local time resolution is found to be insufficient in the validation set, the pooling window can be reduced to 1×2 to balance the amount of computation and the degree of feature retention. During this period, the loss value of each training round is monitored and compared with the pre-given target loss threshold, which is determined empirically after a compromise between speech recognition accuracy and model complexity. If the loss value remains above the threshold in 10 consecutive training rounds, the upper limit of the training round is increased and the parameter update is repeated. When the loss value is lower than the threshold and remains stable in several additional training rounds, the model training is considered complete. After the last training, the recognition accuracy is verified with the test set, and the time series information and spectrum information output by the convolutional layer and the pooling layer are distinguished. The feature data of each moment and the corresponding frequency band are respectively summarized and recorded to form time series and spectrum separation data that can be used for subsequent analysis.

[0023] Based on the time series and spectrum separation data, first determine the dimensional range of the input layer and merge the time dimension index and frequency component label into a multidimensional vector. Then, when using the multilayer perceptron, specify the number of hidden layers and neurons. For example, arrange 128 neurons in the first hidden layer and 64 neurons in the second hidden layer. If signs of overfitting are observed in the training process, grab a number of samples from the previously sorted validation set for cross-checking and introduce a dropout rate between 0.2 and 0.3 in the corresponding hidden layer. Then set the loss function to cross entropy and use the back-propagation method based on stochastic gradient descent to update the weights and biases. After each update, compare the training loss with the validation loss. Difference. If the training loss shows a downward trend while the validation loss continues to rise in multiple iterations, the learning rate is reduced to 1 / 10 of the original value and the overfitting tendency is observed in subsequent iterations. In order to confirm the recognition accuracy of the model in different time periods and frequency bands, segmented statistics are performed on several batches of data and compared with the pre-set accuracy reference value. For example, the range of 0 to 0.8 is considered acceptable, and values ​​greater than 0.8 are considered too high and additional fine-tuning of the hidden layer capacity is required. Finally, when the training and validation process is completed, the important correlation weights between each time point and the corresponding frequency component are analyzed and recorded. The key time-frequency spectrum features are extracted through the multi-layer perceptron to generate a time-frequency spectrum feature vector.

[0024] The steps to obtain the speech dynamic analysis model are: The time-frequency spectrum feature vector is input into the gated recurrent unit, and the initial sequence analysis is performed on the input data, which is converted into a serialized format to generate serialized feature data; Based on the serialized feature data, the intonation and rhythm in the speech data are learned and analyzed, and the feature processing process is iterated through loop connections to adjust the data representation and build a speech dynamic analysis model.

[0025] Specifically, when the time-frequency eigenvector is input into the gated recurrent unit, the length and number of channels of the input sequence are first set according to the dimension range of the eigenvector obtained previously. Then, the thresholds of the update gate and the reset gate are defined in the gating process. The training data is scanned multiple times to determine whether these thresholds need to be adjusted. For example, the threshold of the update gate can be initially set to 0.5 and the difference between the update gate output and the historical hidden state is observed in subsequent iterations. If there is still a large error after 10 iterations, the threshold is lowered to 0.45 and trained again. If the error shows a downward trend, the threshold is maintained and the trend of the loss value is recorded. In order to ensure that some key time points are not missed when performing the initial sequence analysis on the input data, In order to obtain the characteristic information of the segment, the output of each time step needs to be compared with the input state of the next time step and the sequence consistency index needs to be calculated. The index is compared with the pre-established reference range. For example, if the consistency index is in the range of 0 to 0.6, it is judged to be in the medium range. If it is higher than 0.6, the memory factor weight inside the gated unit can be appropriately increased in subsequent processing. At the same time, in order to make the sequence analysis produce a stable serialization format, the stability of the time step output under different speech segments can be tested after each round of iterative training. The difference between time steps greater than 0.2 is considered to be not stable enough and then recalculated. Finally, a stable sequence output format is obtained and recorded in the process of multiple rounds of iterations and continuous comparison to generate serialized feature data.

[0026] When learning and analyzing the intonation and rhythm in speech data based on serialized feature data, the number of layers of recurrent connections can be set to 2 to 3 and the hidden state dimensions of each layer can be distinguished. For example, the hidden state size of the first recurrent layer is set to 64 and the second recurrent layer is set to 32. By gradually iterating feature processing on the training set and recording the changes in loss values ​​at different time steps, it is confirmed whether the recurrent layers need to be increased or decreased. If the training loss fluctuates and cannot be reduced to the predetermined range during multiple iterations, check whether the parameter update rate between the layers deviates from the previously recorded value range. If the deviation is large, reduce the learning step size within the recurrent unit. And compare the changing trend of the loss value again. In order to make the differences in intonation and rhythm better reflected in the feature representation, the serialized feature data can be additionally divided into pitch component and rhythm component in each layer for separate tracking. If it is found that the rhythm component increases abnormally in the later stage of training, the training rounds are increased by 10% and re-verified in combination with the update gate and reset gate thresholds obtained previously. During this period, the gate thresholds are fine-tuned in steps of 0.1. If the error curve is stable after several consecutive iterations, this state is retained to generate the final output data. After completing all the cyclic connections and iterative feature processing processes, the data representation is adjusted to build a speech dynamic analysis model.

[0027] The steps to obtain the user authentication result are: Use the speech dynamic analysis model to process the time-frequency spectrum feature vector, identify the user's speech pattern, and generate preliminary authentication data; Based on the preliminary identity verification data, statistical analysis is performed to calculate the identity verification rate. The calculation formula is: ; in, represents the identification rate, is the user feature vector elements, is the first feature vector The average value of the elements, For the The standard deviation of the elements, represents the contribution rate of each element to identity recognition, and m represents the total number of elements in the feature vector; Based on the identity recognition rate, the degree of match with the preset standard is evaluated. If the identity recognition rate exceeds the preset standard, the user's identity is confirmed, otherwise access is denied and the user identity verification result is obtained.

[0028] Specifically, when using the speech dynamic analysis model to process the time-frequency spectrum feature vector and identify the user's speech pattern, it is necessary to first read the previously obtained time-frequency spectrum feature vector and arrange it in the form of continuous frames. For the data points of each frame, the pre-counted user pitch range is used as a reference, and the acoustic energy concentration is compared with the previously obtained speech feature energy range. When comparing, frames with energy lower than 0.01 or higher than 2.5 are classified as abnormal ranges and checked. During the process, the frequency of pitch changes is continuously recorded and the recorded results are compared with the historical average. For example, the pitch is compared with the 80Hz to 300Hz range, the speech rate is compared with the 2 to 7 syllables per second range, and the pronunciation characteristics are compared with the sound quality. The features are mapped with the timbre parameters obtained previously, and the continuity and stability of the voice pattern are confirmed through multiple cycles, and the time periods with large deviations are marked. When the recorded deviations exceed the reference threshold, the voiceprint components of the specific time period are re-checked and the integrity of the vocal tract spectrum is observed. The reference threshold can be determined according to the variation range of the pitch and speaking speed in the common voice fragments of the user collected previously. After multiple rounds of comparison and cumulative analysis, the judgment data for distinguishing abnormal pronunciation from normal pronunciation is obtained. Through these judgment data, the basic patterns of the user's voice features can be identified and classified as an overall feature. Finally, the feature output is used as the preliminary data format for identity authentication and recorded to generate preliminary identity authentication data.

[0029] The benefit of the formula is that it combines the user feature vector with the mean and standard deviation, and introduces the contribution of each element. Item, to achieve comprehensive consideration of the degree of difference and the weight of difference impact; The acquisition step is to obtain the first The values ​​of the elements are read out and normalized according to the acoustic parameters recorded in the training set. The acquisition step is to count the long-term average values ​​of the corresponding elements in all homologous speech feature vectors and perform floating point precision correction. The steps for obtaining are to calculate the variance of the corresponding elements in the same speech feature vector set and then take the square root to obtain the standard deviation. The acquisition step is to quantitatively evaluate the contribution of the voice element to identity recognition and record the actual distribution value between 0 and 1, and determine its relative importance in combination with the user voice diversity parameter collected previously; Calculation process: When When , , , , calculate the numerator first: ; Then calculate the denominator: ; So the first half of the fraction is: ; Then calculate The cumulative product of , , , so: ; ; ; Multiply the three terms together: ; The final overall result is The result shows that the identity recognition rate is 1.08, which can be compared with the recognition threshold defined previously. The larger the value, the higher the consistency with the user's original characteristics.

[0030] When evaluating the degree of match with the preset standard based on the identity recognition rate, it is necessary to first organize the identity recognition rate sequence obtained during the continuous monitoring process and compare it with the historical standard sequence. By recording and summarizing the recognition rate of each day or each voice interaction, a judgment range is set according to the average user recognition rate and fluctuation range collected previously. For example, a recognition rate lower than 0.5 is considered to be obviously abnormal and requires further comparison of component features. A recognition rate between 0.5 and 1.0 is considered to be within an acceptable range. If it exceeds 1.0, it is classified as a high match group. If the recognition rate is higher than 1.0 for multiple consecutive times, it is marked as an extremely high match. Each detection is compared with the user benchmark distribution calculated previously. Yes, if the cumulative recognition rate is lower than 0.5 for multiple times, continue to read the sound source and check the pitch, energy concentration and speaking speed to observe whether the current user has any changes in sound state or other voice abnormalities. The process includes arranging the daily average and time period peak values ​​of the aforementioned recognition rate records and matching them within the corresponding numerical range. If there is a deviation, mark the moment when the deviation value exceeds 0.1, and then review it during the next interaction. Finally, after these comparison operations are completed, the stability of the user recognition results is obtained and the final mark is confirmed. Through this process, it can be determined whether the preset standard is met and the user is immediately allowed to pass when the value exceeds the standard, otherwise access is denied, and the user identity verification result is obtained.

[0031] The steps to obtain abnormal data identification results are: Based on the user authentication result, the sensors of the smart wearable device are activated to collect physiological and environmental data and generate real-time sensor data streams; Based on the real-time sensor data stream, the isolation forest algorithm is used for anomaly detection to calculate the anomaly index of each data point. The calculation formula is: ; in, Represents data points The abnormal indicators It is a data point In the The eigenvalue of dimension, is the importance coefficient of the dimension feature, is the characteristic dimension; Based on the abnormal index, the degree of deviation of each data point from the normal range is evaluated, and the data points with abnormal index higher than the threshold are identified as abnormal, generating abnormal data identification results.

[0032] Specifically, when activating the sensor of the smart wearable device based on the user identity authentication result, it is necessary to first read whether the previously recorded identity authentication result has valid authorization information, and then check the preset sampling frequency and number of sampling channels before turning on the sensor. The sampling frequency can be set within the range of 5 to 10 times per second based on the detection statistics of environmental noise and the power consumption of the equipment. If it is detected that the sampling frequency still deviates from the set range after multiple tests, the working mode inside the sensor is inspected and the inspection information is recorded, and then the appropriate physiological data and environmental data monitoring category is selected. For example, in terms of physiological data, multiple items such as heart rate, body temperature, and blood oxygen saturation are divided for synchronous acquisition, and in terms of environmental data, items such as temperature, humidity, and concentrations of certain chemical components in the air are divided for parallel collection, and the above-collected physiological and environmental data are collected. The data are arranged in chronological order and indexed. If there is data missing or abnormal jumps during the process, check the current power consumption and connection stability of the sensor and collect data for the corresponding period again. At the same time, the sensor's self-test log is recorded during the entire collection process and compared with the established sensor accuracy reference value. For example, the current real-time collected heart rate is compared with the previously established 40 to 180 times per minute range, the body temperature is compared with the 30℃ to 45℃ range, the air humidity is compared with the 10% to 90% range, and the harmful gas concentration is compared with the 0ppm to 10ppm range. If a monitoring value exceeds these ranges continuously, further troubleshooting is carried out or the results of multiple rounds of repeated sampling are observed. When all are met, the multi-channel collection results are packaged and summarized and updated to the real-time sensor data stream to generate a real-time sensor data stream.

[0033] The formula is useful in that The comprehensive weights of multi-dimensional features are accumulated and used The function maps the result to the interval (0, 1) to quantify the degree of abnormality of a single data point; The acquisition step is to extract the physiological data and environmental data collected previously in the first Dimensions and data points The corresponding measured values ​​are converted and normalized. The acquisition step is to record the statistical probability of abnormality of the same dimensional features in the past period of time, and then set the weight value according to the probability and the historical accuracy of the sensor. The acquisition step is to determine the total number of multi-dimensional features used and uniquely number each feature component; Calculation process: If ,make , where 0.25 may come from the normalized value of a physiological indicator (such as body temperature), 0.60 may come from the normalized result of an environmental indicator (such as humidity), and 0.10 may come from the normalized result of another physiological or environmental indicator. These parameters are obtained by calculating the abnormal distribution probability of multiple sets of sensor data and weighted averaging. The weights will fluctuate in the range of 0.1 to 0.6 depending on the distribution of monitoring data and equipment differences. First calculate the index part: ; Then plug in the index: ; ; Add 1 to get the denominator: ; thereby: ; The results show that a data point The abnormal index is 0.585. If judged in combination with the predetermined standard, when the value is greater than 0.50, it is classified as abnormal for observation; if it is less than 0.50, it is classified as normal range.

[0034] When evaluating the degree of deviation from the normal range for each data point based on abnormal indicators, it is necessary to first compare the sampling values ​​of each dimension with the benchmark interval, and compare the values ​​of different indicators such as heart rate, body temperature, humidity, etc. obtained by statistics with their corresponding reasonable intervals frame by frame, and then summarize the abnormal indicators corresponding to each frame and record them in a time series. In order to ensure the continuity of the entire judgment process, it is necessary to update the abnormal mark of the current frame after each iterative comparison. By superimposing the abnormal mark on the time series and retrieving those time periods where abnormalities appear continuously in multiple frames, it is observed whether it is a short-term fluctuation or a continuous deviation. If these marked values ​​are higher than the specified reference threshold for multiple consecutive times, they will be tracked. For example, setting a reference The threshold is in the range of about 0.50. The threshold is set by collecting and statistically analyzing the deviation of sensor data in similar environmental scenarios in the past. If the abnormal index rises rapidly to above 0.80 in a short period of time, it is regarded as an obvious abnormal situation. Then the heart rate or harmful gas concentration components in this period are compared to see whether they exceed their normal ranges at the same time to prove the validity of the sensor readings. If multiple indicators are near the high threshold or have crossed the threshold, the corresponding period will be characterized as an abnormal point and a timestamp will be attached. These abnormal points will be identified and listed in a unified manner during subsequent analysis. At the end of the entire process, a judgment list indicating whether the data point is abnormal or not will be obtained, and the abnormal data identification result will be generated.

[0035] The steps to obtain the automatic correction data results are: Based on the abnormal data identification results, the Kalman filter algorithm is applied to identify abnormal data points and separate them from the data stream of the smart wearable device to generate a data stream that has been preliminarily cleaned; Based on the data stream that has been initially cleaned, the smoothing value of each data point is calculated using the following formula: ; in, Representative data points The smoothed value of and Represents data points Data points before and after, and is the weight factor, Represents the range of data points; Based on the smoothed value, the data stream is readjusted to obtain the automatically corrected data result.

[0036] Specifically, when the abnormal data points are identified based on the abnormal data identification results and the Kalman filter algorithm is applied, it is necessary to first read which data points are marked as abnormal from the abnormal data identification results obtained previously, and then locate these abnormal data points and corresponding timestamps in the Kalman filter processing stage. In the process, the variance of the sensor measurement noise should be compared with the historical measurement error data, and the prediction and update steps in the filtering process should be derived through the sensor state transition relationship recorded multiple times. At the same time, each data point should be compared and analyzed. If it deviates significantly from the statistical mean interval or remains deviant after repeated measurements, it is compared again to confirm whether it is indeed abnormal, and the confirmed abnormal points are removed. In this process, it is also necessary to pay attention to whether the accuracy and sampling frequency of the device itself are within the established range. For example, the sampling frequency should be maintained between 5 and 10 times per second, and if it deviates from this range, the credibility of the filter needs to be re-evaluated. After confirmation, normal points are distinguished from abnormal points and abnormal points are eliminated, and all normal data are re-aggregated in chronological order to obtain a data stream that has been preliminarily cleaned.

[0037] The formula is useful because it uses the The weighted values ​​of neighboring data points are used to balance the data points. interference factors, and with the help of the denominator Achieve a balance of weighted results; The acquisition steps are to first determine the data points Index in the time series and extract the previous and next data points, and The acquisition step is to read the real sampling value of the adjacent position in the data stream after preliminary cleaning. and The acquisition step is to refer to the past statistical results of the smoothness of adjacent points and analyze the deviation distribution accumulated by multiple rounds of continuous sampling of the device, and assign different weights to these adjacent points. The acquisition step is to select the number of neighboring points according to the previously recorded sampling rate and the required smoothness, which is generally between 3 and 5, and may be adjusted accordingly according to the characteristics of the device for different scenarios; Calculation process: make , select data points At the adjacent points at the sequence position index, it is assumed that they are observed and recorded by the device: ; ; The weight factor is determined based on the mean square error of the last 100 samples and the floating range of adjacent values. If the following weights are obtained: ; Each of these They are all in the range of 0.2 to 0.5 and may be fine-tuned as the data fluctuates, but here we take fixed values ​​as an example and calculate the left and right sums in turn: ; ; Add these two together: ; Divide by : ; The results show that at this moment the data point The smoothing value is 0.3503. If it is compared with the average water level value of the equipment collected before, it is about 1.00, which means that the data point has deviated at this time. In the subsequent smoothing correction, the deviation will be gradually returned to the nearby benchmark value.

[0038] When the data stream is readjusted based on the smoothing value, it is necessary to first continuously retrieve the smoothing value of each data point, arrange the smoothing values ​​corresponding to all data points in chronological order and compare them with the original data point values. If the difference in the smoothing values ​​of some data points between adjacent frames exceeds the set uniform transition threshold, they are re-segmented and the amplitude change of the data segment is further observed. If the amplitude continues to deviate from the reference range, it is compared with the sensor noise characteristics collected previously to see if there is a periodic jitter cause. For each analysis, the final difference mean and variance can be statistically calculated through multiple iterations of smoothing operations and compared with thresholds such as 0.1 or 0.2. The threshold is set by the device in actual operation to record the sensor drift rate and output instability. If multiple observations are within the range, this batch of data will be included in the available list. If it is still exceeded, it will prompt a subsequent inspection to see if it is caused by sensor failure or sudden change of environment. Finally, the smoothing correction results of each time period are globally numbered and summarized into a complete and smoothly connected data sequence to obtain the automatic correction data results.

[0039] The steps to obtain the device health status results are as follows: Apply adaptive filtering to automatically correct the data results and obtain an optimized data stream; Based on the optimized data flow, by comparing the performance between data segmentation sets, the operating status of the smart wearable device is analyzed and evaluated, the health status of the device is determined, and the device health status results are generated.

[0040] Specifically, when applying adaptive filtering to process automatic correction data results, it is necessary to first distinguish the time series of continuous sampling in the automatic correction data results obtained previously, and select the dynamic characteristics of tracking noise as the basis for adjusting the adaptive filtering. During this period, an error reference value for measuring the amplitude of noise change can be established for each observation window. For example, the mean square error range from the most recent 50 samples is included in the calculation and the reference value is updated in real time during the comparison process. If the mean square error in several windows deviates from the reference value by more than a specified amplitude, the filter parameters are adjusted. In specific implementation, each data point will be compared with the historical observation value and the filter gain will be increased or decreased in combination with the error signal. If the trend of the error signal is still detected to deviate from the preset threshold interval after multiple increases, it is necessary to re-evaluate the sampling environment and observe the sensor working mode. The reference threshold interval It usually comes from the error distribution formed by multiple sampling of the same device at normal temperature and pressure. By calculating the distribution mean and discreteness, an upper and lower limit is obtained and recorded for real-time adjustment. When the error signal is lower than the upper and lower limits for multiple consecutive frames, it means that the current filter gain is appropriate. If it is not satisfied, the filter weight coefficient is updated and iterated again. During the whole process, the time series generated after filtering will be compared with the previous unfiltered sequence in segments. For example, each segment of data is divided into several small groups and compared to see whether there is a significant deviation at the same time point. If a single point deviation is found to exceed the predetermined standard, the original data will be further tracked to see whether there is sensor noise drift or sudden abnormality. After completing the filtering correction of all data segments, the filtered sequences are spliced ​​together to obtain a time series data with higher smoothness and lower noise impact, thus obtaining an optimized data stream.

[0041] Based on the optimized data stream and by comparing the performance between data segmentation sets, it is necessary to first divide the optimized data stream into several subsets according to different time periods or different load states, and then record the statistical characteristics of each subset that are consistent with the normal operating indicators of the equipment. For example, in physiological sensor measurements, data such as heart rate or body temperature can be grouped and their range of changes in the time dimension can be counted. The average value of each group is compared with the historical normal value. If it is found that the average value of multiple samples deviates from the set range in a short period of time, the phenomenon is recorded and the data details are continued to be traced. The performance comparison between data segmentation sets can adopt multiple comparison methods, such as quantitative analysis of the peak value, valley value and change rate in each subset, and the situation where the change rate exceeds the reference value of 0.2 or 0.3 is marked. The reference value The value is obtained from the historical volatility distribution collected earlier, and will increase or decrease a certain proportion according to the usage scenario of the device. Then the annotation results of each subset are integrated and arranged. If it is found that some subsets have abnormal frequencies that are too high or the vibration amplitude is significantly increased, it is further inquired whether there is a correlation with the external environment. A similar method can also be used when cross-checking indicators such as pressure, temperature, current, and voltage. Each indicator has a specific reasonable range, such as pressure comparison 0MPa to 2MPa, current comparison 0A to 5A, etc. Each cycle comparison is to monitor whether it crosses this range and the number of times it crosses. Finally, the distribution and fluctuation characteristics of different indicators between subsets are combined to analyze the overall operation status of the smart wearable device and make corresponding judgments to generate equipment health status results.

[0042] An identity authentication system, comprising: The voiceprint collection module receives voice samples from users, extracts time series and spectrum information from them, and generates time-spectrum feature vectors; The voiceprint analysis module inputs the time spectrum feature vector into the sequence processing unit, analyzes the intonation and rhythm, and generates a speech dynamic analysis model; The identity verification module uses the voice dynamic analysis model to compare the stored user model, verify the voice consistency, and obtain the user identity verification result; The data anomaly detection module analyzes the sensor data stream of the smart wearable device based on the user identity authentication result, identifies abnormal data points, and generates abnormal data identification results; The data correction and health assessment module adjusts and corrects the data output based on the abnormal data identification results, applies the verification process to the automatically corrected data, and obtains the equipment health status results.

[0043] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A biometric-based smart wearable device identity authentication method, characterized in that: The following steps are involved: Collect user voice samples, remove background noise from the voice samples using spectral subtraction and Wiener filter technology to generate denoised voice samples; use convolutional neural network to separate the time series and spectrum information of the voice signal from the denoised voice samples, and extract the time spectrum feature vector; Input the time-frequency spectrum feature vector into the gated recurrent unit, strengthen the learning of intonation and rhythm in the speech data through sequence analysis, and generate a speech dynamic analysis model; verify the user identity through the speech dynamic analysis model to obtain a user identity verification result; Based on the user identity authentication result, the sensor data stream of the smart wearable device is collected, and the isolation forest algorithm is used to perform independence analysis on each data point in the data stream, identify the data points that deviate from the normal range, and generate an abnormal data identification result; Based on the abnormal data identification result, adjusting the data output according to the normal data, correcting the abnormal data, and outputting the automatic correction data result; Adaptive filtering is applied to the automatic correction data results, and the automatic correction data results are verified using cross-validation to obtain equipment health status results.

2. The biometric-based smart wearable device identity authentication method according to claim 1, characterized in that: The steps of obtaining the denoised speech sample are as follows: Collect user speech samples, apply frequency domain analysis through spectral subtraction, identify interference from non-speech background noise, and generate preliminary denoised speech samples; Based on the preliminary denoised speech sample, a Wiener filter is used for processing, and the speech signal is optimized by spectrum reconstruction to obtain a further denoised speech sample; Based on the further denoised speech sample, sound quality detection and adjustment are performed, and signal quality evaluation is performed to obtain a denoised speech sample.

3. The biometric-based smart wearable device identity authentication method according to claim 1, characterized in that: The steps of obtaining the time-frequency spectrum feature vector are as follows: Performing deep feature analysis on the denoised speech sample, processing the speech signal using a convolutional neural network, separating the time series and spectrum information in the speech signal through a convolution layer and a pooling layer, and obtaining time series and spectrum separation data; Based on the time series and spectrum separation data, a multi-layer perceptron is used to analyze and identify the relationship between each time point and the frequency component, extract key time spectrum features, and generate a time spectrum feature vector.

4. The biometric-based smart wearable device identity authentication method according to claim 1, characterized in that: The steps for obtaining the speech dynamic analysis model are as follows: Input the time-frequency spectrum feature vector into a gated recurrent unit, perform initial sequence analysis on the input data, convert it into a serialized format, and generate serialized feature data; Based on the serialized feature data, the intonation and rhythm in the speech data are learned and analyzed, and the feature processing process is repeatedly iterated through loop connections to adjust the data representation and build a speech dynamic analysis model.

5. The biometric-based smart wearable device identity authentication method according to claim 1, characterized in that: The steps for obtaining the user identity authentication result are: Using the speech dynamic analysis model to process the time spectrum feature vector, identify the user's speech pattern, and generate preliminary identity authentication data; Based on the preliminary identity verification data, statistical analysis is performed to calculate the identity verification rate, and the calculation formula is: ； in, represents the identification rate, is the user feature vector elements, is the first feature vector The average value of the elements, For the The standard deviation of the elements, represents the contribution rate of each element to identity recognition, and m represents the total number of elements in the feature vector; Based on the identity recognition rate, the degree of matching with the preset standard is evaluated. If the identity recognition rate exceeds the preset standard, the user identity is confirmed, otherwise access is denied, and the user identity verification result is obtained.

6. The biometric-based smart wearable device identity authentication method according to claim 1, characterized in that: The steps for obtaining the abnormal data identification result are: Based on the user identity authentication result, activate the sensor of the smart wearable device, collect physiological and environmental data, and generate a real-time sensor data stream; Based on the real-time sensor data stream, the isolation forest algorithm is applied to perform anomaly detection and calculate the anomaly index of each data point. The calculation formula is: ； in, Represents data points The abnormal indicators It is a data point In the The eigenvalue of dimension, is the importance coefficient of the dimension feature, is the characteristic dimension; Based on the abnormal index, the degree of deviation of each data point from the normal range is evaluated, and the data points with abnormal indexes higher than the threshold are identified as abnormal, and an abnormal data identification result is generated.

7. The biometric-based smart wearable device identity authentication method according to claim 1, characterized in that: The steps of obtaining the automatic correction data result are: Based on the abnormal data identification result, a Kalman filter algorithm is applied to identify abnormal data points, and the abnormal data points are separated from the data stream of the smart wearable device to generate a data stream that has been preliminarily cleaned; Based on the preliminarily cleaned data stream, the smoothing value of each data point is calculated using the following formula: ； in, Representative data points The smoothed value of and Represents data points Data points before and after, and is the weight factor, Represents the range of data points; Based on the smoothed value, the data stream is readjusted to obtain an automatically corrected data result.

8. The biometric-based smart wearable device identity authentication method according to claim 1, characterized in that: The steps for obtaining the device health status result are: Applying adaptive filtering to process the automatic correction data results to obtain an optimized data stream; Based on the optimized data stream, by comparing the performance between the data segmentation sets, the operating status of the smart wearable device is analyzed and evaluated, the health status of the device is determined, and the device health status result is generated.

9. An identity authentication system for a smart wearable device identity authentication method based on biometrics according to any one of claims 1 to 8, characterized in that: include: The voiceprint collection module receives voice samples from users, extracts time series and spectrum information from them, and generates time-spectrum feature vectors; The voiceprint analysis module inputs the time spectrum feature vector into the sequence processing unit, analyzes the intonation and rhythm, and generates a speech dynamic analysis model; The identity verification module uses the voice dynamic analysis model to compare the stored user model, verify the voice consistency, and obtain the user identity verification result; The data anomaly detection module analyzes the sensor data stream of the smart wearable device based on the user identity authentication result, identifies abnormal data points, and generates abnormal data identification results; The data correction and health assessment module adjusts and corrects the data output based on the abnormal data identification results, applies the verification process to the automatically corrected data, and obtains the equipment health status results.

Images