Non-contact infusion monitoring method based on radio frequency acquisition system

Abstract

The invention discloses a non-contact infusion monitoring method based on a radio frequency acquisition system, and relates to the field of infusion monitoring. The problem that CSI data are unstable due to the fact that a wireless channel is interfered by a multipath effect, environment noise and equipment drift is effectively solved. According to the method, a channel physical model is constructed, a linear relation between the liquid level height and the phase difference is established based on the model, approximate linear mapping is further formed, a first estimated value of the liquid remaining amount is solved in combination with target overall channel state information, and effective modeling of a channel change causal mechanism is achieved. In order to further improve the monitoring accuracy, a linear regression model based on a feature vector is also introduced to obtain a second estimated value, and a third estimated value is obtained through the trained deep reasoning network. The three are integrated to obtain a target estimation value, and early warning judgment is performed in combination with the coordinates of the infusion bottle and model output, so that the early warning accuracy is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of infusion monitoring, and more particularly to a non-contact infusion monitoring method based on a radio frequency acquisition system. Background Technology

[0002] In traditional clinical intravenous infusion care, monitoring of the infusion status mainly relies on manual rounds or the addition of gravity / pressure sensors to the infusion tubing. While contact-based sensing solutions can achieve basic monitoring of drip rate and remaining medication, they have several limitations: gravity sensors are extremely sensitive to changes in patient position and limb movement, easily triggering false alarms due to slight traction; they also rely on battery power, resulting in limited battery life and the need for regular maintenance, affecting the stability of long-term continuous monitoring. Furthermore, the shortage of nursing staff in primary healthcare institutions makes it difficult to maintain high-frequency manual rounds, leading to the failure to promptly detect abnormal events such as infusion termination and leakage, increasing medical risks and impacting nursing efficiency.

[0003] With the development of the Internet of Things (IoT) and wireless sensing technologies, non-contact infusion monitoring has become a research hotspot. In recent years, WiFi Channel State Information (CSI), due to its high-dimensional characteristics and penetrating propagation properties, has been used for sensing fluid level and leakage status, demonstrating advantages such as non-contact, low cost, and scalable deployment. The Angle of Arrival (AoA) positioning technology introduced by the Bluetooth 5.1 standard, combined with Mesh networks, has also provided a foundation for ward-level device networking and centimeter-level positioning. However, practical applications still face multiple challenges: wireless channels are susceptible to multipath effects, environmental noise, and device drift interference, leading to unstable raw CSI data; the lack of effective modeling of the causal mechanisms of channel changes limits sensing accuracy. Existing systems have not yet formed a complete closed loop from signal preprocessing and edge intelligent discrimination to alarm push, making it difficult to provide stable, accurate, and low-false-alarm intelligent infusion monitoring solutions for clinical use. Therefore, a novel non-contact infusion monitoring method integrating radio frequency sensing, edge artificial intelligence, and low-power communication is urgently needed to improve the safety, efficiency, and patient experience of the infusion process. Summary of the Invention

[0004] To achieve accurate monitoring of non-contact infusion, this invention proposes a non-contact infusion monitoring method based on a radio frequency (RF) acquisition system. The RF acquisition system includes an edge calculator and multiple lighting-RF fusion nodes communicatively connected to it. These nodes are installed in lighting fixtures or lighting circuits within the ward and coupled to the lighting power supply system for continuous power supply. Each node integrates an RF transceiver module and an antenna assembly, with the antenna assembly radiating towards the ground. The method includes the following steps performed by the edge calculator:

[0005] The RF transceiver module in the lighting-RF fusion node periodically transmits probe signals, and after each probe, it synchronously acquires the corresponding overall channel state information based on the orthogonal frequency division multiplexing modulation mechanism.

[0006] Based on the influence of the dielectric target consisting of the infusion tube, dripping bucket and liquid surface on the detection signal, i.e. the equivalent single-path fading increment caused by the change in liquid level, a channel physical model is constructed.

[0007] A linear model between liquid level height and phase difference is constructed based on the channel physical model; the phase difference represents the phase change of the same subcarrier at adjacent time points; the linear model corresponds one-to-one with the subcarrier.

[0008] Phase unwrapping and common-mode drift removal are performed sequentially on the acquired overall channel state information;

[0009] The missing or outlier values ​​in the overall channel state information after removing drift are filled in to obtain the target overall channel state information.

[0010] Construct the feature vector corresponding to each time step based on the overall channel state information of the target.

[0011] An approximate linear mapping relationship between liquid level and phase difference is constructed using a linear model. This approximate linear mapping relationship, along with the overall channel state information of the target, is used to calculate the remaining liquid volume in the infusion bottle as a first estimate. The remaining liquid volume in the infusion bottle is then calculated as a second estimate by inputting the feature vector into a linear regression benchmark model.

[0012] A training set is constructed, and a pre-defined deep inference network is trained using the training set and a loss function to obtain the target infusion monitoring model. The training set includes multiple training samples. Each training sample consists of a feature vector and a training label corresponding to a time step. The training label includes: the actual remaining volume of the infusion bottle, the actual drip rate, and the infusion status at the corresponding time step.

[0013] Using the constructed feature vector as model input, the remaining amount of liquid in the infusion bottle is predicted by the target infusion monitoring model to obtain a third estimate.

[0014] The target estimate is obtained by solving the first, second, and third estimates, and the coordinates of the infusion bottle are calculated. The early warning is then issued based on the target estimate, the output of the target infusion monitoring model, and the coordinates of the infusion bottle.

[0015] Furthermore, the overall channel state information includes channel state information corresponding to each subcarrier; the channel state information includes at least: the signal amplitude value and phase value corresponding to the corresponding subcarrier, wherein the amplitude value reflects the attenuation degree of the signal on the corresponding subcarrier, and the phase value reflects the phase offset caused by the signal propagation path.

[0016] Furthermore, the formula expression for the overall channel state information is as follows:

[0017]

[0018] In the formula, k represents the discrete time point at which sampling is performed at a preset time interval, i.e., the time step; i represents the subcarrier index number, N s Indicates the number of subcarriers; φ k,i H represents the original phase value of the i-th subcarrier at time point k; k,i | represents the signal amplitude value of the i-th subcarrier at time point k; j is the imaginary unit, j 2 =-1; H k (f i ) represents the complex form of the channel frequency response of the i-th subcarrier, i.e., the i-th subcarrier at time k and frequency f. i Channel state information;

[0019] The formula for the channel physical model is as follows:

[0020]

[0021] In the formula, L is the effective multipath number, representing the number of different paths the signal travels through; a l,k τ represents the signal amplitude value of the l-th path at time point k; f represents the frequency variable; τ l,k Δh represents the time delay of the l-th path at time point k; k H represents the equivalent single-path decay increment caused by changes in liquid level height. k (f) represents the overall channel state information at time point k and frequency f;

[0022] The formula for the linear model between the liquid level height and the phase difference is (each subcarrier can be calculated individually for a corresponding Δφ). k ):

[0023]

[0024] In the formula, h k Indicates the liquid level height at time point k; n liq λ represents the refractive index of the drug solution; λ0 represents the approximate wavelength of all subcarriers; ε φ,k Indicates the noise term; Δφ k This indicates the change in phase gradient caused by the drop in liquid level.

[0025] Furthermore, the acquisition of overall channel state information is sequentially subjected to phase unwrapping and common-mode drift removal, specifically as follows:

[0026] Phase unwrapping: based on φ k,i The main domain ∈(-π,π] is used to identify phase transition points that cross the ±π main domain in the overall channel state information and record them as δ. k,p And the true continuous phase value is obtained through cumulative compensation:

[0027] In the formula, δ k,p The flag variable represents whether the p-th subcarrier undergoes a ±π phase transition at time point k: if it is a transition from -π to π, then δ is set. k,p =+1, if it is a jump from π to -π, then set δ. k,p =-1, or set δ if there is no transition. k,p =0; This represents the unwrapped phase value of the i-th subcarrier at time point k; This represents the accumulation of all positions in the first i subcarriers where ±π transitions have occurred;

[0028] Common-mode drift removal: Based on the unwrapped continuous phase values, the least squares iterative method is used to estimate and remove the common-mode drift term caused by the frequency shift, obtaining the corrected phase, specifically including:

[0029] Based on the continuous phase value after untangling Construct a linear model for subcarrier index i: In the formula, β k γk represents the global phase shift coefficient at time point k; i Let be the linear phase drift coefficient of the i-th subcarrier at time point k, and represent the slope of the change with subcarrier index i; ∈ k,i This represents the modeling residual term, reflecting the detection noise;

[0030] Using a linear model of subcarrier index i to analyze the drift parameter β k and γ k Perform an initial estimate to minimize the sum of squared errors: In the formula, They represent β respectively k With γ k The corresponding estimated value;

[0031] Through estimated values The corrected phase value is obtained using the following formula:

[0032] In the formula, This represents the phase value of the i-th subcarrier at time point k after removing common-mode drift.

[0033] Furthermore, the process of filling in missing or outlier values ​​in the overall channel state information after removing drift to obtain the target overall channel state information specifically involves:

[0034] A first-order random walk model is established based on the overall channel state information after drift removal; the formula for the first-order random walk model is: In the formula, This represents the i-th subcarrier at time point k and frequency f. i Channel state information after drift removal; w k-1 This represents process noise, i.e., the random changes in the channel state between time points k-1 and k; the first-order random walk model is used to determine the channel state based on the previous time step. Predict the overall channel state information after drift removal at time k.

[0035] Based on the detection noise v k Construct a correction formula; the modified formula is used to correct the overall channel state information predicted by the first-order random walk model. The corrected formula is: In the formula, v k Indicates the detection noise, z k express The corresponding correction value;

[0036] Using the prior noise covariance matrix and the observation noise covariance, the standard Kalman filter recursive algorithm is executed based on a first-order random walk model and a modified formula to obtain the i-th subcarrier at time point k and frequency f. i Channel state information estimation value;

[0037] For missing or outlier values ​​in the data, Akima interpolation is used to fill in the missing or outlier values ​​based on the estimated channel state information, and finally the overall channel state information of the target is obtained.

[0038] Furthermore, the feature vector corresponding to each time step includes the mean and standard deviation of the channel state information amplitude, the mean and standard deviation of the phase, and the first m fast Fourier transform spectral coefficients of the phase difference sequence calculated at the corresponding time step, which are used to capture the long-term trend of the liquid level.

[0039] The formula expression for the feature vector is:

[0040]

[0041] In the formula, μ represents the mean, |H| represents the amplitude in the overall channel state information of the target, and σ represents the standard deviation; This represents the phase value in the overall channel state information of the target; and Let $k$ represent the mean and standard deviation of the amplitude at time point $k$, respectively. and These represent the mean and standard deviation of the phase at time point k, respectively; This represents the first m Fast Fourier Transform spectral coefficients of the phase difference sequence from time point k-τ to k, used to capture the long-term trend of the liquid level.

[0042] Furthermore, the process of constructing an approximate linear mapping relationship between liquid level height and phase difference using a linear model, and then using this approximate linear mapping relationship to determine the remaining liquid volume in the infusion bottle as a first estimated value, specifically involves:

[0043] The phase difference change of all subcarriers at each time step is averaged using a linear model between liquid level height and phase difference. Specifically, the linear model between liquid level height and phase difference is transformed as follows: Based on the modified formula, a formula is constructed to average the phase difference changes of all subcarriers at each time step: In the formula, This represents the phase difference of the i-th subcarrier at time point k in the overall channel state information of the target.

[0044] An approximate linear mapping relationship between liquid level and phase difference is constructed based on the formula for averaging; the remaining liquid volume in the infusion bottle is then calculated using this approximate linear mapping relationship as the first estimated value.

[0045] The expression for the approximately linear mapping relationship between the liquid level height and the phase difference is: c′ k =C0-ρh k In the formula, C0 represents the total amount of liquid in the infusion bottle, and ρ represents the conversion factor between the cross-sectional area and density of the infusion bottle; c′ k This represents the first estimate at time point k.

[0046] Furthermore, a second estimate of the vector is obtained by inputting the feature vector into the linear regression benchmark model, where the formula expression of the linear regression benchmark model is:

[0047] c″ k =w T F k +b; where c″ k This represents the second estimate at time point k; w represents the weight vector, which consists of model parameters, and its dimension is the same as that of the feature vector F. k Same; b indicates the bias term.

[0048] Furthermore, the formula for the loss function is as follows:

[0049] In the formula, σc c represents the learnable log-variance, used to control the weights for the residual estimation task; k This represents the actual remaining amount of the infusion bottle at time point k. This represents the third estimate obtained from the model prediction; This represents the L1 norm loss in the residual estimate;

[0050] σ v V represents the learnable log-variance, used to control the weights for the drip rate estimation task; k This represents the actual drip rate of the infusion bottle at time point k; This indicates the predicted drip rate from the model; This represents the L1 norm loss of the drip rate estimation;

[0051] σ s s represents the learnable log-variance, used to control the weights for the infusion state classification task; k This indicates the actual infusion state of the infusion bottle at time point k; This indicates the infusion status predicted by the model; This represents the cross-entropy loss function, used to measure the error in classification tasks; This represents the loss value.

[0052] Furthermore, the radio frequency acquisition system also includes:

[0053] Three known location-based base stations are connected to the edge calculator for communication; each of the base stations is equipped with an 8-element UWB-BLE composite antenna.

[0054] The positioning base station is used to receive a time-stamped signal emitted from the illumination-RF fusion node and reflected back by the infusion bottle through an 8-element UWB-BLE composite antenna, calculate the corresponding angle of arrival based on the signal, and send it to the edge calculator.

[0055] The calculation of the infusion bottle coordinates specifically involves determining the coordinates of the infusion bottle based on the angle of arrival calculated from the three positioning base stations.

[0056] Compared with the prior art, the present invention has at least the following beneficial effects:

[0057] (1) This invention effectively solves the problem of unstable CSI data caused by multipath effects, environmental noise, and equipment drift interference in wireless channels by sequentially performing phase unwrapping and common-mode drift removal on the acquired overall channel state information. By constructing a channel physical model and establishing a linear relationship between liquid level and phase difference based on this model, an approximate linear mapping is further formed. Combined with the overall channel state information of the target, the first estimate of the remaining liquid is obtained, realizing effective modeling of the causal mechanism of channel changes and improving the estimation accuracy. To further improve the monitoring accuracy, the system also introduces a linear regression model based on feature vectors to obtain a second estimate, and obtains a third estimate through a trained deep inference network. The target estimate is obtained by combining the three estimates, and the warning judgment is made by combining the coordinates of the infusion bottle and the model output, which significantly improves the accuracy and reliability of the warning and enhances the practicality and intelligence level of this invention.

[0058] (2) This invention embeds a radio frequency sensing module into the lighting fixtures or wiring in the ward, and utilizes indoor WiFi channel status information (CSI) to achieve non-contact monitoring of the remaining fluid volume and abnormal conditions during the infusion process. Compared with traditional sensors that require direct contact with the infusion tubing, this non-contact method avoids false alarms caused by patient activity, reduces interference with patients, and improves patient comfort.

[0059] (3) The present invention uses three positioning base stations with known locations, equipped with an 8-element UWB-BLE composite antenna, to receive signals emitted from the illumination-RF fusion node and reflected back by the infusion bottle, calculate the direction of arrival (AoA), and determine the specific location of the infusion bottle through the edge calculator, thereby achieving precise positioning of the infusion bottle.

[0060] (4) In this invention, the lighting-RF fusion node is installed in the lighting fixtures or lighting circuits inside the ward and coupled to the lighting power supply system. This solves the problem of limited battery life of traditional battery-powered equipment, eliminates the risk of equipment downtime, ensures stable operation for a long time, and eliminates the need for frequent maintenance.

[0061] (5) By sequentially performing phase unwrapping and common-mode drift removal on the acquired overall channel state information, and completing the missing or outlier values, the overall channel state information of the target is obtained. Based on this, the feature vectors containing the mean and standard deviation of the amplitude, the mean and standard deviation of the phase of the channel state information calculated at the corresponding time step, and the first m fast Fourier transform spectral coefficients of the phase difference sequence are constructed. The target infusion monitoring model trained by the deep inference network can accurately predict the remaining liquid volume and infusion status in the infusion bottle, reducing the false alarm rate and improving the monitoring accuracy.

[0062] (6) The method of the present invention is applicable to medical institutions of all sizes, especially primary hospitals. It can effectively monitor the infusion process even when manpower is tight, and prevent medical accidents caused by negligence. At the same time, due to its contactless design and low cost, it is easy to promote to large-scale wards, and has good economic efficiency and practicality. Attached Figure Description

[0063] Figure 1 This is a flowchart of a non-contact infusion monitoring method based on a radio frequency acquisition system in an embodiment of the present invention. Detailed Implementation

[0064] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0065] To achieve accurate monitoring of non-contact infusion, such as Figure 1 As shown, this invention proposes a non-contact infusion monitoring method based on a radio frequency (RF) acquisition system. The RF acquisition system includes an edge calculator and multiple lighting-RF fusion nodes communicatively connected to it. These nodes are installed in lighting fixtures or lighting circuits within the ward and coupled to the lighting power supply system for continuous power supply. Each node integrates an RF transceiver module and an antenna assembly, with the antenna assembly radiating towards the ground. The method includes the following steps performed by the edge calculator:

[0066] The RF transceiver module in the lighting-RF fusion node periodically transmits probe signals, and after each probe, it synchronously acquires the corresponding overall channel state information based on the orthogonal frequency division multiplexing modulation mechanism.

[0067] The overall channel state information includes channel state information corresponding to each subcarrier; the channel state information includes at least: the signal amplitude value and phase value corresponding to the corresponding subcarrier, wherein the amplitude value reflects the attenuation degree of the signal on the corresponding subcarrier, and the phase value reflects the phase offset caused by the signal propagation path.

[0068] The formula for the overall channel state information is as follows:

[0069]

[0070] In the formula, k represents the discrete time point at which sampling is performed at a preset time interval, i.e., the time step; i represents the subcarrier index number, N s Indicates the number of subcarriers; φ k,i H represents the original phase value of the i-th subcarrier at time point k; k,i | represents the signal amplitude value of the i-th subcarrier at time point k; j is the imaginary unit, j 2 =-1; Hk (f i ) represents the complex form of the channel frequency response of the i-th subcarrier, i.e., the i-th subcarrier at time k and frequency f. i Channel state information;

[0071] Based on the influence of the dielectric target consisting of the infusion tube, dripping bucket and liquid surface on the detection signal, i.e. the equivalent single-path fading increment caused by the change in liquid level, a channel physical model is constructed.

[0072] The formula for the channel physical model is as follows:

[0073]

[0074] In the formula, L is the effective multipath number, representing the number of different paths the signal travels through; a l,k τ represents the signal amplitude value of the l-th path at time point k; f represents the frequency variable; τ l,k Δh represents the time delay of the l-th path at time point k; k H represents the equivalent single-path decay increment caused by changes in liquid level height. k (F) represents the overall channel state information at time point k and frequency f;

[0075] A linear model between liquid level height and phase difference is constructed based on the channel physical model; the phase difference represents the phase change of the same subcarrier at adjacent time points; the linear model corresponds one-to-one with the subcarrier.

[0076] The formula for the linear model between the liquid level height and the phase difference is (each subcarrier can be calculated individually for a corresponding Δφ). k ):

[0077]

[0078] In the formula, h k Indicates the liquid level height at time point k; n liq λ represents the refractive index of the drug solution; λ0 represents the approximate wavelength of all subcarriers; ε φ,k Indicates the noise term; Δφ k This indicates the change in phase gradient caused by the drop in liquid level.

[0079] Phase unwrapping and common-mode drift removal are performed sequentially on the acquired overall channel state information;

[0080] The process of sequentially performing phase unwrapping and common-mode drift removal on the acquired overall channel state information is as follows:

[0081] Phase unwrapping: based on φ k,iThe main domain ∈(-π,π] is used to identify phase transition points that cross the ±π main domain in the overall channel state information and record them as δ. k,p And the true continuous phase value is obtained through cumulative compensation:

[0082] In the formula, δ k,p The flag variable represents whether the p-th subcarrier undergoes a ±π phase transition at time point k: if it is a transition from -π to π, then δ is set. k,p =+1, if it is a jump from π to -π, then set δ. k,p =-1, or set δ if there is no transition. k,p =0; This represents the unwrapped phase value of the i-th subcarrier at time point k; This represents the accumulation of all positions in the first i subcarriers where ±π transitions have occurred;

[0083] Common-mode drift removal: Based on the unwrapped continuous phase values, the least squares iterative method is used to estimate and remove the common-mode drift term caused by the frequency shift, obtaining the corrected phase, specifically including:

[0084] Based on the continuous phase value after untangling Construct a linear model for subcarrier index i: In the formula, β k γk represents the global phase shift coefficient at time point k; i Let be the linear phase drift coefficient of the i-th subcarrier at time point k, and represent the slope of the change with subcarrier index i; ∈ k,i This represents the modeling residual term, reflecting the detection noise;

[0085] Using a linear model of subcarrier index i to analyze the drift parameter β k and γ k Perform an initial estimate to minimize the sum of squared errors: In the formula, They represent β respectively k With γ k The corresponding estimated value;

[0086] Through estimated values The corrected phase value is obtained using the following formula:

[0087] In the formula, This represents the phase value of the i-th subcarrier at time point k after removing common-mode drift.

[0088] In this invention, common-mode drift can also be removed using the following formula:

[0089]

[0090] In the formula, a0 and a1 represent the global phase shift coefficient and the linear phase drift coefficient, respectively, which are obtained by least squares fitting.

[0091] The missing or outlier values ​​in the overall channel state information after removing drift are filled in to obtain the target overall channel state information.

[0092] The process of filling in missing or outlier values ​​in the overall channel state information after removing drift to obtain the target overall channel state information specifically involves:

[0093] A first-order random walk model is established based on the overall channel state information after drift removal; the formula for the first-order random walk model is: In the formula, This represents the i-th subcarrier at time point k and frequency f. i Channel state information after drift removal; w k-1 This represents process noise, i.e., the random changes in the channel state between time points k-1 and k; the first-order random walk model is used to determine the channel state based on the previous time step. Predict the overall channel state information after drift removal at time k.

[0094] Based on the detection noise v k Construct a correction formula; the modified formula is used to correct the overall channel state information predicted by the first-order random walk model. The corrected formula is: In the formula, v k Indicates the detection noise, z k express The corresponding correction value;

[0095] Using the prior noise covariance matrix and the observation noise covariance, the standard Kalman filter recursive algorithm is executed based on a first-order random walk model and a modified formula to obtain the i-th subcarrier at time point k and frequency f. i Channel state information estimation value;

[0096] For missing or outlier values ​​in the data, Akima interpolation is used to fill in the missing or outlier values ​​based on the estimated channel state information, and finally the overall channel state information of the target is obtained.

[0097] Construct the feature vector corresponding to each time step based on the overall channel state information of the target.

[0098] The feature vector corresponding to each time step includes the mean and standard deviation of the amplitude, the mean and standard deviation of the phase of the channel state information calculated at the corresponding time step, and the first m fast Fourier transform spectral coefficients of the phase difference sequence.

[0099] The formula expression for the feature vector is:

[0100]

[0101] In the formula, μ represents the mean, |H| represents the amplitude in the overall channel state information of the target, and σ represents the standard deviation; This represents the phase value in the overall channel state information of the target; and Let $k$ represent the mean and standard deviation of the amplitude at time point $k$, respectively. and These represent the mean and standard deviation of the phase at time point k, respectively; This represents the first m Fast Fourier Transform spectral coefficients of the phase difference sequence from time point k-τ to k, used to capture the long-term trend of the liquid level.

[0102] An approximate linear mapping relationship between liquid level and phase difference is constructed by using a linear model between the liquid level and phase difference. The remaining amount of liquid in the infusion bottle is then estimated using this approximate linear mapping relationship and the overall channel state information of the target.

[0103] The process involves constructing an approximate linear mapping relationship between liquid level and phase difference using a linear model, and then using this approximate linear mapping relationship to determine the remaining liquid volume in the infusion bottle as a first estimate. Specifically:

[0104] The phase difference change of all subcarriers at each time step is averaged using a linear model between liquid level height and phase difference. Specifically, the linear model between liquid level height and phase difference is transformed as follows: Based on the modified formula, a formula is constructed to average the phase difference changes of all subcarriers at each time step: In the formula, This represents the phase difference of the i-th subcarrier at time point k in the overall channel state information of the target.

[0105] An approximate linear mapping relationship between liquid level and phase difference is constructed based on the formula for averaging; the remaining liquid volume in the infusion bottle is then calculated using this approximate linear mapping relationship as the first estimated value.

[0106] The expression for the approximately linear mapping relationship between the liquid level height and the phase difference is: c′ k =C0-ρh k In the formula, C0 represents the total amount of liquid in the infusion bottle, and ρ represents the conversion factor between the cross-sectional area and density of the infusion bottle; c′ k This represents the first estimate at time point k.

[0107] The second estimate is obtained by inputting the feature vector into the linear regression benchmark model to solve for the remaining liquid volume in the infusion bottle.

[0108] The second estimate of the vector is obtained by inputting the feature vector into the linear regression benchmark model, where the formula expression of the linear regression benchmark model is:

[0109] c″ k =w T F k +b; where c″ k This represents the second estimate at time point k; w represents the weight vector, which consists of model parameters, and its dimension is the same as that of the feature vector F. k Same; b indicates the bias term.

[0110] A training set is constructed, and a pre-defined deep inference network is trained using the training set and a loss function to obtain the target infusion monitoring model. The training set includes multiple training samples. Each training sample consists of a feature vector and a training label corresponding to a time step. The training label includes: the actual remaining volume of the infusion bottle, the actual drip rate, and the infusion status at the corresponding time step.

[0111] In this embodiment, the output of the target infusion monitoring model includes:

[0112] The remaining volume prediction value is the third estimate, the drip rate prediction value, and the infusion status prediction value. Among them: the infusion status prediction value s k ∈{0,1,2}, where 0,1,2 represent normal, low liquid level, and blockage, respectively.

[0113] The target infusion monitoring model comprises the following components connected in sequence:

[0114] 1. Input layer: Accepts feature vector F k These feature vectors are concatenated through channels and then fed into the subsequent convolution stack.

[0115] 2. Convolutional coding layers (ConvBlocks):

[0116] It contains 3 convolutional units (Conv-BN-ReLU-SE), and the specific structure of each unit is as follows:

[0117] Conv: Convolutional layer with a kernel size of 3×3 and output channels of 32, 64 and 128 respectively.

[0118] BN: Batch Normalization.

[0119] ReLU: Activation function ReLU (Rectified Linear Unit).

[0120] SE: The SE (Squeeze-and-Excitation) module is used to enhance feature representations.

[0121] 3. Flattening layer: The feature map of the last layer of the convolutional coding layer is flattened into a one-dimensional vector and used as the input of the subsequent fully connected layer.

[0122] 4. Three parallel task branches:

[0123] The network is divided into three independent and parallel task branches, each used for a different task:

[0124] Regression-C: A regression task used to predict the remaining volume in an IV bottle. It consists of a fully connected layer + a Swish activation function + another fully connected layer, which are connected in sequence, and finally output a real value.

[0125] Regression-V: A regression task used to predict the drip rate of an infusion bottle. It consists of a fully connected layer + a Swish activation function + another fully connected layer, which are connected in sequence, and finally output a real value.

[0126] Classification-S: A classification task used to classify infusion status (e.g., normal, low fluid level, obstruction). It consists of a fully connected layer, a Dropout layer, and a Softmax layer connected in sequence, ultimately outputting a class label 's'. k ∈{0,1,2}.

[0127] These three task branches are structurally parallel, meaning they each obtain features from the flattened layer and process them independently.

[0128] The formula for the loss function is as follows:

[0129]

[0130] In the formula, σ c c represents the learnable log-variance, used to control the weights for the residual estimation task; k This represents the actual remaining amount of the infusion bottle at time point k. This represents the third estimate obtained from the model prediction; This represents the L1 norm loss in the residual estimate;

[0131] σ v V represents the learnable log-variance, used to control the weights for the drip rate estimation task; l This represents the actual drip rate of the infusion bottle at time point k; This indicates the predicted drip rate from the model; This represents the L1 norm loss of the drip rate estimation;

[0132] σ s s represents the learnable log-variance, used to control the weights for the infusion state classification task; k This indicates the actual infusion state of the infusion bottle at time point k; This indicates the infusion status predicted by the model; This represents the cross-entropy loss function, used to measure the error in classification tasks; This represents the loss value.

[0133] (σ c +σ v +σ s This part is the regularization term, which ensures that the weights of each task do not become too large, preventing overfitting. It directly adds the three learnable log-variances as an additional penalty term.

[0134] This embodiment uses the AdamW optimizer for training with an initial learning rate of 1e-3, and adjusts the learning rate using a Cosine-Decay scheduling strategy. During training, the batch size is set to 64, and the total number of iterations is 80 epochs. To enhance the model's robustness, online data augmentation methods include randomly blocking 10% of the subcarriers and applying ±5dB jitter to the RSSI value.

[0135] For model compression and deployment, the convolutional channels are first sorted based on the γ values ​​in the Batch Normalization (BN) layer, and the 40% of channels with the lowest amplitude are pruned to achieve model slimming. Subsequently, the Kullback-Leibler divergence (KLD) calibration method is used to generate a quantization table using 512 batch samples on an NVIDIA A100 GPU, thereby ensuring efficient deployment and operation of the model on edge devices.

[0136] Through the above training strategies and model compression techniques, this invention not only improves the generalization ability of the model, but also significantly reduces the demand for computing resources, enabling the system to run stably and efficiently in real medical environments.

[0137] Using the constructed feature vector as model input, the remaining amount of liquid in the infusion bottle is predicted by the target infusion monitoring model to obtain a third estimate.

[0138] The target estimate is obtained by solving the first, second, and third estimates, and the coordinates of the infusion bottle are calculated. An early warning is issued based on the target estimate, the output of the target infusion monitoring model, and the coordinates of the infusion bottle.

[0139] By sequentially performing phase unwrapping and common-mode drift removal on the acquired overall channel state information, and filling in missing or outlier values, the overall channel state information of the target is obtained. Based on this, an eigenvector is constructed containing the mean and standard deviation of the amplitude, the mean and standard deviation of the phase, and the first m fast Fourier transform spectral coefficients of the phase difference sequence of the channel state information calculated at the corresponding time step. The target infusion monitoring model trained by a deep inference network can accurately predict the remaining liquid volume and infusion status in the infusion bottle, reducing the false alarm rate and improving monitoring accuracy.

[0140] In this embodiment, an early warning is issued based on the target estimate, the output of the target infusion monitoring model, and the coordinates of the infusion bottle. Specifically:

[0141] 1. Normal state:

[0142] When s k =0 and The system starts accumulating time when the current infusion rate is within the normal range.

[0143] 2. Warning status:

[0144] If detected 3 times in a row Or a third estimate This indicates that an abnormal situation is about to occur. In this case, nursing staff can prepare in advance.

[0145] 3. Alarm Status:

[0146] If s k If the state is ∈{1,2} (indicating an abnormal state), or if the warning state continues for more than 30 seconds, an alarm message will be generated: And log it. The alarm information includes the infusion bottle device ID, timestamp t, and current remaining fluid volume. and state s k The data is then transmitted to the nurse's workstation and mobile terminal via a Bluetooth Mesh network for timely notification to relevant personnel. The device ID for the infusion bottle is derived from the bottle's coordinates.

[0147] The radio frequency acquisition system also includes:

[0148] Three known location-based base stations are connected to the edge calculator for communication; each of the base stations is equipped with an 8-element UWB-BLE composite antenna.

[0149] The positioning base station is used to receive a time-stamped signal emitted from the illumination-RF fusion node and reflected back by the infusion bottle through an 8-element UWB-BLE composite antenna, calculate the corresponding angle of arrival based on the signal, and send it to the edge calculator.

[0150] It should be noted that the lighting-RF fusion node both transmits signals and receives overall channel state information. Simultaneously, the positioning base station also receives signals emitted from the lighting-RF fusion node and reflected back from the IV bottle. To ensure the system accurately understands the correspondence between these signals, some synchronization and identification mechanisms are typically required.

[0151] Time synchronization mechanism: All nodes (including lighting-RF fusion nodes and base stations) need a common time reference. This can be achieved through the Network Time Protocol (NTP) or a more precise clock synchronization mechanism (such as IEEE 1588PTP).

[0152] Timestamp: Each transmitted probe signal is timestamped. When the signal is reflected back and received, the receiving node records the timestamp of the received signal. By comparing the transmitted and received timestamps, it can be determined which signals are responses to the same probe signal.

[0153] The method for calculating the coordinates of the infusion bottle is as follows: the coordinates of the infusion bottle are determined based on the angle of arrival calculated from three positioning base stations. Specifically, the steps include:

[0154] 1. Calculate the angular error Δθ introduced by BLE frequency offset compensation using the temperature-RSSI-frequency offset ternary correction model:

[0155] Δθ=α T (T-25℃)+α P (RSSI+60)+α F Δf; where:

[0156] Δf represents the carrier frequency drift, which is estimated using the DCR-PLL algorithm;

[0157] α T ,α P ,α F This represents an empirical coefficient used to correct for the effects of temperature, RSSI (Received Signal Strength Indication), and frequency drift.

[0158] T represents ambient temperature; RSSI represents received signal strength indication.

[0159] 2. Correct the arrival direction angle:

[0160] Based on the Δθ calculated using the above model, the arrival direction angle θ of the original measurement is... mn Make corrections:

[0161] θ′ mn =θ mn -Δθ;

[0162] In the formula, θ mnLet θ represent the direction angle (AoA) of arrival of the infusion bottle relative to two base stations m and n, where m ≠ n and m, n ∈ {1, 2, 3}. Specifically, θ mn The angle is calculated based on the signals received by base stations m and n. Δθ represents the angular error caused by factors such as temperature, RSSI (Received Signal Strength Indication), and frequency drift. This error is calculated using a temperature-RSSI-frequency drift ternary correction model. θ′ mn This represents the corrected direction of arrival angle, used for subsequent position calculation and positioning processing.

[0163] 3. Use the corrected arrival direction angle θ′ mn To calculate the angular relationship between the infusion bottle and the two base stations m and n:

[0164]

[0165] In the formula, m ≠ n and m, n ∈ {1, 2, 3};

[0166] θ′ mn This represents the calibrated direction of arrival angle of the infusion bottle relative to base station m and base station n;

[0167] p m This represents the location coordinates of the m-th base station;

[0168] p n This represents the location coordinates of the nth base station;

[0169] p represents the coordinates of the infusion bottle at an unknown location;

[0170] • Represents the dot product (inner product) of vectors.

[0171] 4. Solve for the coordinates of the infusion bottle using the Law of Cosines:

[0172] Assume the coordinates of the three positioning base stations are p1, p2, and p3, and the coordinates of the infusion bottle at the unknown location are p = [x, y, z]. T :

[0173] p1, p2, p3: represent the three-dimensional coordinates of the three positioning base stations, respectively;

[0174] p: represents the coordinates of the infusion bottle at an unknown location, which is a three-dimensional vector;

[0175] x, y, z: represent the coordinate components of the infusion bottle in three-dimensional space.

[0176] Calculate the distance relationship from the infusion bottle to each base station using the law of cosines:

[0177]

[0178] In the formula:

[0179] ||pp m ||: Indicates the distance between the infusion bottle and the m-th base station;

[0180] ||p||: represents the modulus of the infusion bottle's coordinates (i.e., the distance from the infusion bottle to the origin);

[0181] ||p m ||: represents the modulus of the coordinates of the m-th base station (i.e., the distance from the base station to the origin);

[0182] d m : Indicates the distance between the infusion bottle and the m-th base station.

[0183] It should be noted that:

[0184] Global coordinate system: In positioning and spatial calculations, a global coordinate system is usually set. For example, in a ward or hospital environment, a corner of the room can be selected as the origin (0,0,0) of the global coordinate system, and the X-axis, Y-axis and Z-axis can represent the width, length and height of the room, respectively.

[0185] Base station location coordinates: The location coordinates p1, p2, p3 of the three known base stations are defined relative to this global coordinate system. This means that the position of each base station is determined relative to the origin of the global coordinate system.

[0186] Infusion bottle position coordinates: The coordinates of an infusion bottle at an unknown location are also defined relative to this global coordinate system.

[0187] 5. Perform joint linearization on the distance and angle relationships of the three sets of base stations (m,n), and use the weighted least squares method to solve for the estimated position of the infusion bottle:

[0188]

[0189] In the formula:

[0190] This represents the estimated position of the infusion bottle obtained by weighted least squares method;

[0191] arg min: indicates the parameter value that minimizes the objective function;

[0192] w m : Represents the weighting coefficient corresponding to the m-th base station, used to reflect the reliability of measurement data from different base stations;

[0193] (||pp m ||-d m ) 2: Represents the squared difference between the actual distance and the measured distance between the infusion bottle and the m-th base station, used to measure the error.

[0194] This formula represents estimating the position of the infusion bottle by minimizing the objective function. The objective function is the sum of squared distance errors from the infusion bottle to the three base stations, and each error term is multiplied by a corresponding weight w. m This is a nonlinear optimization problem, which usually requires iterative algorithms (such as gradient descent) to solve.

[0195] In this way, the present invention first corrects the arrival direction angle, and then uses the corrected arrival direction angle to calculate the position. This ensures that various possible error factors are taken into account when calculating the position of the infusion bottle, thereby improving the accuracy of positioning.

[0196] This invention effectively solves the problem of unstable CSI data caused by multipath effects, environmental noise, and equipment drift interference in wireless channels by sequentially performing phase unwrapping and common-mode drift removal on the acquired overall channel state information. By constructing a channel physical model and establishing a linear relationship between liquid level and phase difference based on this model, an approximate linear mapping is formed. Combined with the overall channel state information of the target, a first estimate of the remaining liquid volume is obtained, achieving effective modeling of the causal mechanism of channel changes and improving estimation accuracy. To further improve monitoring accuracy, the system also introduces a linear regression model based on feature vectors to obtain a second estimate, and a third estimate is obtained through a trained deep inference network. The target estimate is obtained by combining these three estimates, and a warning judgment is made by combining the infusion bottle coordinates and model output, significantly improving the accuracy and reliability of the warning and enhancing the practicality and intelligence level of this invention.

[0197] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0198] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0199] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0200] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A non-contact infusion monitoring method based on a radio frequency acquisition system, characterized in that, The radio frequency acquisition system includes: an edge calculator, and multiple lighting-RF fusion nodes communicatively connected to it. These nodes are installed in lighting fixtures or lighting circuits within the ward and coupled to the lighting power supply system to achieve continuous power supply. Each node integrates a radio frequency transceiver module and an antenna assembly, with the antenna assembly radiating towards the ground. The method includes the following steps performed by the edge calculator: The RF transceiver module in the lighting-RF fusion node periodically transmits probe signals, and after each probe, it synchronously acquires the corresponding overall channel state information based on the orthogonal frequency division multiplexing modulation mechanism. Based on the influence of the dielectric target consisting of the infusion tube, dripping bucket and liquid surface on the detection signal, i.e. the equivalent single-path fading increment caused by the change in liquid level, a channel physical model is constructed. A linear model between liquid level height and phase difference is constructed based on the channel physical model; the phase difference represents the phase change of the same subcarrier at adjacent time points; the linear model corresponds one-to-one with the subcarrier. Phase unwrapping and common-mode drift removal are performed sequentially on the acquired overall channel state information; The missing or outlier values ​​in the overall channel state information after removing drift are filled in to obtain the target overall channel state information. Construct the feature vector corresponding to each time step based on the overall channel state information of the target. An approximate linear mapping relationship between liquid level and phase difference is constructed using a linear model. This approximate linear mapping relationship, along with the overall channel state information of the target, is used to calculate the remaining liquid volume in the infusion bottle as a first estimate. The remaining liquid volume in the infusion bottle is then calculated as a second estimate by inputting the feature vector into a linear regression benchmark model. A training set is constructed, and a pre-defined deep inference network is trained using the training set and a loss function to obtain the target infusion monitoring model. The training set includes multiple training samples. Each training sample consists of a feature vector and a training label corresponding to a time step. The training label includes: the actual remaining volume of the infusion bottle, the actual drip rate, and the infusion status at the corresponding time step. Using the constructed feature vector as model input, the remaining amount of liquid in the infusion bottle is predicted by the target infusion monitoring model to obtain a third estimate. The target estimate is obtained by solving the first, second, and third estimates, and the coordinates of the infusion bottle are calculated. The early warning is then issued based on the target estimate, the output of the target infusion monitoring model, and the coordinates of the infusion bottle.

2. The non-contact infusion monitoring method based on a radio frequency acquisition system according to claim 1, characterized in that, The overall channel state information includes the channel state information corresponding to each subcarrier; The channel state information includes at least: the signal amplitude value and phase value corresponding to the corresponding subcarrier, wherein the amplitude value reflects the degree of signal attenuation on the corresponding subcarrier, and the phase value reflects the phase shift caused by the signal propagation path.

3. The non-contact infusion monitoring method based on a radio frequency acquisition system according to claim 2, characterized in that, The formula for the overall channel state information is as follows: In the formula, k represents the discrete time point at which sampling is performed at a preset time interval, i.e., the time step; i represents the subcarrier index number, N s Indicates the number of subcarriers; φ k,i H represents the original phase value of the i-th subcarrier at time point k; k,i | represents the signal amplitude value of the i-th subcarrier at time point k; j is the imaginary unit, j 2 =-1; H k (f i ) represents the complex form of the channel frequency response of the i-th subcarrier, i.e., the i-th subcarrier at time k and frequency f. i Channel state information; The formula for the channel physical model is as follows: In the formula, L is the effective multipath number, representing the number of different paths the signal travels through; a l,k τ represents the signal amplitude value of the l-th path at time point k; f represents the frequency variable; τ l,k Δh represents the time delay of the l-th path at time point k; k H represents the equivalent single-path decay increment caused by changes in liquid level height. k (f) represents the overall channel state information at time point k and frequency f; The formula for the linear model between the liquid level height and the phase difference is as follows: In the formula, h k Indicates the liquid level height at time point k; n 1iq λ represents the refractive index of the drug solution; λ0 represents the approximate wavelength of all subcarriers; ε φ,k Indicates the noise term; Δφ k This indicates the change in phase gradient caused by the drop in liquid level.

4. The non-contact infusion monitoring method based on a radio frequency acquisition system according to claim 3, characterized in that, The process of sequentially performing phase unwrapping and common-mode drift removal on the acquired overall channel state information is as follows: Phase unwrapping: based on φ k,i The main domain ∈(-π,π] is used to identify phase transition points that cross the ±π main domain in the overall channel state information and record them as δ. k,p And the true continuous phase value is obtained through cumulative compensation: In the formula, δ k,p The flag variable represents whether the p-th subcarrier undergoes a ±π phase transition at time point k: if it is a transition from -π to π, then δ is set. k,p =+1, if it is a jump from π to -π, then set δ. k,p =-1, or set δ if there is no transition. k,p =0; This represents the unwrapped phase value of the i-th subcarrier at time point k; This represents the accumulation of all positions in the first i subcarriers where ±π transitions have occurred; Common-mode drift removal: Based on the unwrapped continuous phase values, the least squares iterative method is used to estimate and remove the common-mode drift term caused by the frequency shift, obtaining the corrected phase, specifically including: Based on the continuous phase value after untangling Construct a linear model for subcarrier index i: In the formula, β k γk represents the global phase shift coefficient at time point k; i Let be the linear phase drift coefficient of the i-th subcarrier at time point k, and represent the slope of the change with subcarrier index i; ∈ k,i This represents the modeling residual term, reflecting the detection noise; Using a linear model of subcarrier index i to analyze the drift parameter β k and γ k Perform an initial estimate to minimize the sum of squared errors: In the formula, They represent β respectively k With γ k The corresponding estimated value; Through estimated values The corrected phase value is obtained using the following formula: In the formula, This represents the phase value of the i-th subcarrier at time point k after removing common-mode drift.

5. A non-contact infusion monitoring method based on a radio frequency acquisition system according to claim 4, characterized in that, The process of filling in missing or outlier values ​​in the overall channel state information after removing drift to obtain the target overall channel state information specifically involves: A first-order random walk model is established based on the overall channel state information after drift removal; the formula for the first-order random walk model is: In the formula, This represents the i-th subcarrier at time point k and frequency f. i Channel state information after drift removal; w k-1 This represents process noise, i.e., the random changes in the channel state between time points k-1 and k; the first-order random walk model is used to determine the channel state based on the previous time step. Predict the overall channel state information after drift removal at time k. Based on the detection noise v k Construct a correction formula; the modified formula is used to correct the overall channel state information predicted by the first-order random walk model. The corrected formula is: In the formula, v k Indicates the detection noise, z k express The corresponding correction value; Using the prior noise covariance matrix and the observation noise covariance, the standard Kalman filter recursive algorithm is executed based on a first-order random walk model and a modified formula to obtain the i-th subcarrier at time point k and frequency f. i Channel state information estimation value; For missing or outlier values ​​in the data, Akima interpolation is used to fill in the missing values ​​based on the channel state information estimates, and finally the overall channel state information of the target is obtained.

6. The non-contact infusion monitoring method based on a radio frequency acquisition system according to claim 5, characterized in that, The feature vector corresponding to each time step includes the mean and standard deviation of the amplitude, the mean and standard deviation of the phase of the channel state information calculated at the corresponding time step, and the first m fast Fourier transform spectral coefficients of the phase difference sequence. The formula expression for the feature vector is: In the formula, μ represents the mean, |H| represents the amplitude in the overall channel state information of the target, and σ represents the standard deviation; This represents the phase value in the overall channel state information of the target; and Let $k$ represent the mean and standard deviation of the amplitude at time point $k$, respectively. and These represent the mean and standard deviation of the phase at time point k, respectively; This represents the first m Fast Fourier Transform spectral coefficients of the phase difference sequence from time point k-τ to k, used to capture the long-term trend of the liquid level.

7. A non-contact infusion monitoring method based on a radio frequency acquisition system according to claim 6, characterized in that, The process involves constructing an approximate linear mapping relationship between liquid level and phase difference using a linear model, and then using this approximate linear mapping relationship to determine the remaining liquid volume in the infusion bottle as a first estimate. Specifically: The phase difference change of all subcarriers at each time step is averaged using a linear model between liquid level height and phase difference. Specifically, the linear model between liquid level height and phase difference is transformed as follows: Based on the modified formula, a formula is constructed to average the phase difference changes of all subcarriers at each time step: In the formula, This represents the phase difference of the i-th subcarrier at time point k in the overall channel state information of the target. An approximate linear mapping relationship between liquid level and phase difference is constructed based on the formula for averaging; the remaining liquid in the infusion bottle is then calculated using this approximate linear mapping relationship as the first estimated value. The expression for the approximately linear mapping relationship between the liquid level height and the phase difference is: c′ k =C0-ρh k In the formula, C0 represents the total amount of liquid in the infusion bottle, and ρ represents the conversion factor between the cross-sectional area and density of the infusion bottle; c′ k This represents the first estimate at time point k.

8. A non-contact infusion monitoring method based on a radio frequency acquisition system according to claim 7, characterized in that, The second estimate of the vector is obtained by inputting the feature vector into the linear regression benchmark model, where the formula expression of the linear regression benchmark model is: c″ k =w T F k +b; where c″ k This represents the second estimate at time point k; w represents the weight vector, which consists of model parameters, and its dimension is the same as that of the feature vector F. k Same; b indicates the bias term.

9. A non-contact infusion monitoring method based on a radio frequency acquisition system according to claim 8, characterized in that, The formula for the loss function is as follows: In the formula, σ c c represents the learnable log-variance, used to control the weights for the residual estimation task; k This represents the actual remaining amount in the infusion bottle at time point k. This represents the third estimate obtained from the model prediction; This represents the L1 norm loss in the residual estimate; σ v V represents the learnable log-variance, used to control the weights for the drip rate estimation task; k This represents the actual drip rate of the infusion bottle at time point k; This indicates the predicted drip rate from the model; This represents the L1 norm loss of the drip rate estimation; σ s s represents the learnable log-variance, used to control the weights for the infusion state classification task; k This indicates the actual infusion state of the infusion bottle at time point k; This indicates the infusion status predicted by the model; This represents the cross-entropy loss function, used to measure the error in classification tasks; This represents the loss value.

10. A non-contact infusion monitoring method based on a radio frequency acquisition system according to claim 1, characterized in that, The radio frequency acquisition system also includes: Three known location-based base stations are connected to the edge calculator for communication; each of the base stations is equipped with an 8-element UWB-BLE composite antenna. The positioning base station is used to receive a time-stamped signal emitted from the illumination-RF fusion node and reflected back by the infusion bottle through an 8-element UWB-BLE composite antenna, calculate the corresponding angle of arrival based on the signal, and send it to the edge calculator. The calculation of the infusion bottle coordinates specifically involves determining the coordinates of the infusion bottle based on the angle of arrival calculated from the three positioning base stations.

Images