A method for self-adaptive control of contact pressure of an ultrasound probe based on respiratory compensation

Abstract

The present application relates to a kind of based on respiratory compensation's ultrasonic probe contact pressure self-adaptive control method, comprising: obtaining body surface fluctuation and contact pressure signal and generating displacement and pressure data;Displacement data is input into respiratory mode prediction model, extracts current respiratory phase feature and infers the expected fluctuation amplitude of next cycle;If amplitude is over threshold, then based on the displacement trend function generated by phase feature, calculate probe position and pressure compensation amount;Accordingly, instruction drives mechanical arm to maintain contact pressure in preset range.The present application realizes the initiative synchronous servo of probe by predicting body surface fluctuation trend, overcomes the lag defect of passive feedback, effectively improves ultrasonic imaging stability and examination efficiency.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic probe contact pressure control technology, and in particular to an adaptive control method for ultrasonic probe contact pressure based on respiratory compensation. Background Technology

[0002] In the field of medical imaging diagnosis, ultrasound examination, with its advantages of no ionizing radiation, real-time dynamic imaging, and relatively low cost, has become an indispensable visualization tool for clinical screening and interventional treatment guidance. However, during actual examinations, the contact state between the ultrasound probe and the patient's body surface is easily affected by the external environment and the patient's own physiological activities. The most typical and unavoidable factor is the patient's respiratory movements. When a patient breathes, the chest and abdominal surface experiences periodic fluctuations. These minute displacement changes are enough to disrupt the original mechanical balance between the ultrasound probe and the skin. If the contact pressure is too low, an air gap will appear between the probe and the skin, causing a large amount of ultrasound waves to be reflected and unable to effectively enter the body, resulting in local acoustic shadowing loss in the image. Conversely, if the contact pressure is too high, it may compress superficial blood vessels or organs, causing deformation of the imaging structure. Therefore, how to enable the ultrasound probe to autonomously adapt to changes in the body surface and maintain a relatively constant contact pressure in the context of the patient's dynamic breathing has become a key technical challenge for improving the efficiency and diagnostic accuracy of ultrasound examinations.

[0003] To address the aforementioned issue of contact instability caused by respiratory movements, current industry practices primarily rely on passive adjustment or manual intervention. One common approach depends on the operator's experience, whereby the physician visually observes changes in screen image quality and adjusts the pressure and angle of the probe based on subjective feel. While this method offers some flexibility, its accuracy is highly dependent on the physician's skill level and fatigue level, and maintaining consistent focus and pressure during prolonged examinations is difficult. Another existing technical solution involves introducing basic force feedback control logic at the robotic arm's end effector. This utilizes a single pressure sensor mounted on the probe's tip to detect the normal contact force. When the detected pressure deviates from a set threshold, a proportional-integral-derivative controller drives the robotic arm to perform rigid pull-back or compression compensation. However, this single-threshold-triggered feedback control strategy suffers from significant technological lag. Since human respiration is a continuous and accelerating process, by the time the pressure sensor detects a significant pressure change, body surface displacement has often already occurred. In this case, the robotic arm's compensation action is essentially a catch-up correction of the predetermined displacement error. This passive response mechanism cannot predict upcoming surface fluctuations, resulting in a persistent dynamic tracking error between the probe and the skin. When the patient takes large movements such as deep breathing, it can easily trigger overshoot and oscillation of the regulatory system, which in turn exacerbates image instability. Summary of the Invention

[0004] Therefore, the technical problem to be solved by this invention is to overcome the shortcomings of existing ultrasound examinations that rely on the operator's manual experience or passive lag feedback based on a single pressure sensor to respond to the patient's respiratory movements. This results in the probe contact pressure being difficult to maintain constant under dynamic fluctuations in the body surface, leading to image flicker artifacts and mechanical adjustment overshoot oscillations. The invention provides an adaptive control method for ultrasound probe contact pressure based on respiratory compensation. By extracting the temporal phase characteristics of the current respiratory movement and pre-inferring the expected body surface fluctuation amplitude in the next cycle based on its intrinsic correlation with historical fluctuation amplitudes, a target compensation amount including position and pressure is actively generated before the actual body surface displacement occurs. This drives the robotic arm to achieve synchronous follow-up application, ultimately stabilizing the contact pressure within a preset range. This effectively improves the continuity and effectiveness of ultrasound image sequences, reduces reliance on operator manual skills, and improves the patient's examination experience.

[0005] To address the aforementioned technical problems, this invention provides an adaptive control method for ultrasonic probe contact pressure based on respiratory compensation, comprising the following steps: The sensor array acquires the surface undulation signal of the tested body and the probe contact pressure signal. After filtering and preprocessing, it generates real-time surface undulation displacement data sequence and contact pressure distribution data. The body surface fluctuation displacement data sequence is input into the pre-built respiratory pattern prediction model. The respiratory pattern prediction model extracts the time phase features that represent the current respiratory movement stage, and based on the inherent relationship between the time phase features and the historical fluctuation amplitude change pattern, it infers and outputs the expected body surface fluctuation amplitude for the next control cycle. In response to the expected body surface fluctuation amplitude exceeding the preset steady-state threshold, the displacement change trend function generated by the breathing pattern prediction model based on the time phase characteristics is used to calculate the target position compensation amount and target pressure regulation amount of the probe used to counteract the respiratory effects. Based on the target position compensation amount and the target pressure adjustment amount, combined with the probe contact pressure distribution data, a sequence of control commands is generated to drive the robotic arm actuator, so that the robotic arm drives the probe to maintain contact with the surface of the object being examined while keeping the contact pressure within a preset range.

[0006] In one embodiment of the present invention, the sensor array includes a flexible piezoresistive thin film sensor and an inertial measurement unit integrated at the front end of the ultrasonic probe. The body surface undulation displacement data sequence is obtained by calculating the acceleration and angular velocity data output by the inertial measurement unit, and the contact pressure distribution data is obtained by analyzing the resistance change array of the flexible piezoresistive thin film sensor.

[0007] In one embodiment of the present invention, the breathing pattern prediction model is constructed using a temporal convolutional network with an encoder-decoder architecture. The temporal convolutional network includes an encoding layer for extracting deep features of the body surface undulation displacement data sequence, and a decoding layer for mapping the deep features to displacement prediction values ​​at future times.

[0008] In one embodiment of the present invention, the training sample set used by the breathing pattern prediction model in the pre-construction stage is constructed by continuously collecting body surface fluctuation displacement data sequences when the patient is in a state of calm breathing, deep breathing and coughing. The label data of the training sample set is the actual body surface displacement value corresponding to a fixed time step after the collection time.

[0009] In one embodiment of the present invention, the extraction process of time phase features specifically includes: In the encoding layer, a dilated convolution operation is performed on the input sequence of body surface displacement data x(t). The output F(t) of the dilated convolution operation at time t satisfies the following relationship: ; Where: Wm is the weight parameter of the convolution kernel at the m position, M is the convolution kernel size, d is the dilation factor, and b is the bias term; The output F(t) of the dilated convolution operation is input into a gated linear unit for nonlinear activation, generating a high-dimensional feature vector H that satisfies the following relationship: ; Where F1(t) and F2(t) are the results of performing a two-channel linear projection on F(t): F1(t) is the information path, carrying specific feature information; F2(t) is the control path, and after being processed by the Sigmoid function σ, the output value is between (0,1); ⊙ represents element-wise multiplication. The generated high-dimensional feature vector H is the temporal phase feature that characterizes the current stage of respiratory movement.

[0010] In one embodiment of the present invention, the step of reasoning and outputting the expected body surface fluctuation amplitude for the next control cycle based on the inherent correlation between time phase characteristics and historical fluctuation amplitude variation patterns specifically includes: In the decoding layer, a weighted summation operation is performed on the high-dimensional feature vector using a fully connected layer. The weight parameters of the fully connected layer are pre-fitted with the slope and extreme values ​​of the body surface displacement changes corresponding to different respiratory phases using a training sample set. Based on the result of the weighted summation operation, a scalar value representing the expected body surface fluctuation amplitude of the next control cycle is directly generated.

[0011] In one embodiment of the present invention, the displacement change trend function is a first-order linear differential function, and the target position compensation amount ΔL of the probe used to counteract the effects of breathing satisfies the following relationship: ; Where t0 is the current time, T is the duration of the next control cycle, and Apred is the expected amplitude of body surface fluctuations. dt is the rate of change of respiratory phase determined based on time phase characteristics, α is a preset displacement compensation proportional factor, and dt is the integral variable.

[0012] In one embodiment of the present invention, the target pressure adjustment amount ΔF satisfies the following relationship: ; Where: β is the preset force-position coupling coefficient; This represents the mean of the pressure distribution data at the current moment.

[0013] In one embodiment of the present invention, the preset steady-state threshold is dynamically adjusted based on the pressure fluctuation variance of multiple sampling points in the contact pressure distribution data within a preset time window. If the pressure fluctuation variance increases, the value of the steady-state threshold is reduced so that the control method is more sensitive to surface fluctuations.

[0014] In one embodiment of the present invention, the step of generating a control command sequence for driving the robotic arm actuator specifically includes: performing inverse kinematics calculation on the target position compensation amount to obtain the target rotation angle of each joint of the robotic arm, and converting the target pressure adjustment amount into impedance control parameters at the end of the robotic arm. The control command sequence includes interpolated trajectory data of the target rotation angle and the impedance control parameters.

[0015] The technical solution of the present invention has the following advantages compared with the prior art: The present invention discloses an adaptive control method for ultrasound probe contact pressure based on respiratory compensation. By acquiring surface fluctuation signals and pressure signals, it predicts the breathing pattern, calculates the compensation amount, generates control commands, and maintains the contact pressure within a preset range. It can predict respiratory fluctuations in real time, actively calculate and compensate for probe position and pressure, thereby stably maintaining the contact pressure between the probe and the subject's surface under dynamic respiratory background, improving the accuracy and reliability of ultrasound imaging, and reducing image jitter and distortion caused by respiratory movements. Attached Figure Description

[0016] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of the steps of the ultrasonic probe contact pressure adaptive control method based on respiratory compensation of the present invention; Figure 2 This is a flowchart of the steps for constructing a respiratory pattern prediction model according to the present invention; Figure 3 This is a flowchart of the steps for extracting temporal phase features based on the respiratory pattern prediction model of the present invention; Figure 4 This is a flowchart of the steps in this invention to infer and output the expected body surface fluctuation amplitude for the next control cycle based on a respiratory pattern prediction model. Figure 5 This is a flowchart illustrating the steps of generating a sequence of control commands for driving a robotic arm actuator according to the present invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0018] Reference Figure 1 As shown, this invention addresses the technical shortcomings of existing technologies, namely, delayed adjustment and lack of predictability, by proposing an adaptive control method for ultrasound probe contact pressure based on respiratory compensation. First, it is necessary to acquire the surface fluctuation signals of the examined body and the probe contact pressure signals collected by a sensor array. After filtering and preprocessing, real-time surface fluctuation displacement data sequences and contact pressure distribution data are generated. This initial step is designed to build a multi-dimensional sensing foundation. Compared to a single force sensor, array-based acquisition not only obtains the average contact pressure under the probe but also captures the displacement waveform of the body surface caused by respiration, providing a continuous and rich time-series data source for subsequent predictive analysis.

[0019] Based on this, the present invention inputs the body surface fluctuation displacement data sequence into a pre-constructed respiratory pattern prediction model. The respiratory pattern prediction model extracts the time phase characteristics representing the current respiratory movement stage, and based on the inherent correlation between the time phase characteristics and the historical fluctuation amplitude variation pattern, it infers and outputs the expected body surface fluctuation amplitude for the next control cycle. The introduction of this technical feature is the core of solving the lag of existing technologies. In principle, human respiratory movement is not a completely random noise signal, but a physiological signal with specific rhythm and envelope characteristics. By analyzing the displacement sequence through the pre-constructed model, it is possible to accurately identify whether the current stage is the early stage of inhalation, the late stage of inhalation, or the expiratory plateau. Only by clarifying the current respiratory phase can the specific amplitude of the body surface bulging or collapsing at the next instant be scientifically inferred based on the motion inertia exhibited by the respiratory phase in historical cycles. This transforms the working mode of the control system from passively sensing and correcting errors to actively predicting trends and intervening in advance.

[0020] Furthermore, in response to the expected fluctuations in body surface area exceeding a preset steady-state threshold, this invention utilizes a respiratory pattern prediction model to generate a displacement change trend function based on time phase characteristics, calculating the target position compensation amount and target pressure adjustment amount of the probe used to counteract the effects of breathing. The technical consideration behind setting the steady-state threshold here is to distinguish between normal physiological micro-movements and effective respiratory displacements requiring interventional compensation, thereby avoiding unnecessary micro-vibrations in the robotic arm when the patient is completely holding their breath or breathing very shallowly, thus contributing to improved system operational stability. Once compensation is determined to be necessary, the system does not simply perform rigid resistance, but rather calculates a set of precise adjustments including spatial position and applied force based on the displacement change trend function.

[0021] Finally, based on the calculated target position compensation and target pressure adjustment, and combined with the probe contact pressure distribution data obtained above, a sequence of control commands is generated to drive the robotic arm actuator, so that the robotic arm can keep the probe in close contact with the surface of the object being examined while maintaining the contact pressure within the preset ideal working range.

[0022] Through the organic combination of the above-mentioned series of technical features, the beneficial effects that the technical solution of the present invention can bring include: Firstly, regarding imaging stability, because the robotic arm performs synchronous follow-up compensation based on precise prediction of respiratory fluctuations, a dynamic fit effect close to soft tracking can be achieved between the probe and the skin. Throughout the patient's entire respiratory cycle, the ultrasound coupling interface remains continuous and pressure is balanced, effectively avoiding image flicker and artifacts caused by probe detachment or excessive pressure, significantly improving the continuity and data validity of ultrasound image sequences.

[0023] Secondly, in terms of automation and efficiency, the method of this invention greatly reduces the reliance on the manual skills of physicians. During robotic-assisted scanning, physicians no longer need to manually counteract the patient's breathing fluctuations to maintain image clarity, and can concentrate more attention on observing and identifying the lesion area. This not only reduces the workload of medical staff, but also makes the examination process more standardized and efficient.

[0024] Finally, in terms of patient experience and safety, the contact pressure maintained by the adaptive algorithm is always below the preset safe and comfortable threshold, avoiding the hard impact of the robotic arm that may be caused by traditional rigid force feedback, and providing the examinee with a more compliant and pressure-free examination experience.

[0025] The method of this invention relies on real-time and accurate acquisition of surface fluctuation signals of the test subject and probe contact pressure signals. In practical applications, how to efficiently and accurately acquire these key signals and convert them into displacement data sequences and pressure distribution data that can be processed later is the foundation for ensuring the effectiveness of the entire adaptive control system.

[0026] In this regard, this application further proposes that the sensor array includes a flexible piezoresistive thin film sensor and an inertial measurement unit integrated at the front end of the ultrasonic probe. The body surface undulation displacement data sequence is obtained by calculating the acceleration and angular velocity data output by the inertial measurement unit, and the contact pressure distribution data is obtained by analyzing the resistance change array of the flexible piezoresistive thin film sensor.

[0027] Specifically, a sensor array is a collection of devices used to sense and acquire physical signals. In this embodiment, the sensor array is designed to simultaneously acquire dynamic fluctuation information of the subject's surface and contact pressure information between the probe and the surface, providing real-time and accurate input data for subsequent respiratory compensation and pressure adaptive control. Integrating the sensor array directly into the front end of the ultrasound probe means that the sensor and the ultrasound probe form a tight integrated unit. This integration ensures that the distance between the sensor and the subject's surface is minimized, thereby improving the directness and accuracy of signal acquisition, reducing measurement errors caused by relative movement or gaps between the probe and the sensor, and facilitating a more compact system design. A flexible piezoresistive thin-film sensor is a pressure sensing element based on the piezoresistive effect, characterized by its excellent flexibility and extremely thin thickness. Its resistance changes predictably when subjected to external pressure. By arranging multiple such sensors in an array, fine sensing of the pressure distribution within the contact area can be achieved. This sensor provides high-resolution pressure distribution data, rather than just the total pressure value, which is crucial for identifying localized high-pressure points and achieving uniform pressure control. An inertial measurement unit (IMU) is an electronic device capable of measuring the acceleration and angular velocity of an object, typically containing an accelerometer and a gyroscope. By processing the acceleration and angular velocity data output by the IMU, the position, attitude, and trajectory of the probe (and the body surface it contacts) in three-dimensional space can be acquired in real time. This enables the system to accurately track the undulating motion of the body surface, providing crucial displacement information for respiratory compensation.

[0028] Furthermore, the body surface undulation displacement data sequence is obtained by processing the acceleration and angular velocity data output by the inertial measurement unit (IMU). The raw acceleration and angular velocity data output by the IMU require complex algorithmic processing to convert into a precise displacement data sequence. This typically involves integrating the acceleration data twice to obtain the displacement, simultaneously using the angular velocity data from the gyroscope for attitude calculation to correct for the gravitational component influence of the accelerometer under different attitudes, and combining techniques such as Kalman filtering or complementary filtering to fuse sensor data, suppress noise and drift, thereby generating a high-precision body surface undulation displacement data sequence. Meanwhile, the contact pressure distribution data is obtained by analyzing the resistance change array of a flexible piezoresistive thin-film sensor. Each sensing unit in the flexible piezoresistive thin-film sensor array changes its resistance value accordingly when subjected to pressure. By measuring the resistance value of each sensing unit in the array and based on a pre-calibrated resistance-pressure conversion relationship, the local pressure borne by each sensing unit can be calculated. Combining these local pressure values ​​forms a two-dimensional distribution data reflecting the contact pressure between the probe and the body surface. This distributed data can provide richer information than a single pressure sensor, helping to identify areas of concentrated pressure and thus enabling more precise pressure regulation.

[0029] In practical applications, human respiratory movements are highly nonlinear and time-varying. Their displacement data sequences often contain complex temporal dependencies and deep features. Traditional prediction models may have difficulty accurately capturing these complex patterns, thus affecting the prediction accuracy of the expected body surface fluctuation amplitude, and consequently affecting the subsequent compensation calculations and the generation of control commands.

[0030] Reference Figure 2 As shown, this application further proposes a breathing pattern prediction model constructed using a temporal convolutional network with an encoder-decoder architecture. This temporal convolutional network is specifically designed to process sequence data with complex temporal dependencies, and its core lies in achieving accurate prediction of breathing patterns through the collaborative work of the encoding and decoding layers.

[0031] Specifically, temporal convolutional networks contain encoding layers for extracting deep features from sequences of body surface undulation displacement data. The role of these encoding layers is to extract abstract, high-level deep features crucial for predicting future breathing patterns from raw, potentially noisy, and complex sequences of body surface undulation displacement data. These deep features better characterize the periodicity, amplitude variation trends, and underlying physiological states of respiratory movements. Encoding layers typically consist of a series of convolutional layers, particularly dilated convolutional layers. Each convolutional layer learns different local patterns of the input sequence through different convolutional kernels, while dilated convolutions allow the receptive field to grow exponentially, capturing longer temporal dependencies without increasing network depth or using pooling operations, thus avoiding information loss. For example, multi-layered dilated convolutional blocks with residual connections can be used to enhance feature extraction capabilities and mitigate the vanishing gradient problem.

[0032] Furthermore, temporal convolutional networks also include decoding layers for mapping deep features to predicted displacement values ​​for future timeframes. The role of this decoding layer is to transform the highly abstract deep features extracted by the encoding layer into concrete, predictable surface displacement values ​​for future timeframes that can be used to guide control. It requires converting these features from high-dimensional space back to low-dimensional, physically meaningful displacement data. The decoding layer can consist of deconvolutional layers or fully connected layers. Deconvolutional layers can progressively recover the length and details of the sequence, while fully connected layers can directly map the feature vector output by the encoder to the expected output dimension, i.e., the predicted displacement value for future timeframes. For example, one or more fully connected layers can be used, taking the final feature vector output by the encoding layer as input, to directly output the expected surface fluctuation amplitude for the next control cycle.

[0033] Specifically, a training sample set is proposed for the pre-construction phase of the breathing pattern prediction model. This set serves as the dataset used to train the temporal convolutional network, with the core objective of enabling the model to learn and recognize different breathing patterns and their evolutionary patterns. This sample set typically contains a large number of input-output pairs, where the input is a sequence of historical body surface displacement data, and the output is the corresponding predicted future body surface displacement. The quality and diversity of the training sample set directly determine the model's generalization ability and prediction accuracy. To ensure that the model can accurately predict body surface fluctuations under various breathing states in practical applications, the training sample set needs to cover as many physiological breathing conditions as possible, thus providing a sufficient learning foundation for the model.

[0034] The construction of the training sample set is specifically accomplished by continuously collecting body surface displacement data sequences while the patient is in a state of calm breathing, deep breathing, and coughing. Calm breathing refers to the normal breathing pattern of the human body in a relaxed state; deep breathing refers to a breathing pattern with significantly increased inhalation and exhalation amplitudes, usually accompanied by large body surface displacements; coughing is a rapid and intense exhalation action, accompanied by severe body surface displacements. By collecting data under these different physiological states, it is ensured that the training sample set can fully reflect the complexity and diversity of human respiratory movements, thereby enabling the respiratory pattern prediction model to learn more robust feature representations and improve its ability to identify and predict various breathing patterns in practical applications. The data acquisition process typically uses high-precision sensors (such as inertial measurement units) to continuously record the displacement changes of the body surface in three-dimensional space.

[0035] The labeled data in the training sample set are the true target values ​​used for supervised learning. For the breathing pattern prediction task, labeled data refers to the actual observed body surface displacement values ​​after a preset fixed time step (e.g., 1 second, 2 seconds, or one control cycle) following the acquisition of the input data sequence (i.e., historical body surface fluctuation displacement data). These true displacement values ​​serve as the model's prediction targets, guiding the model to learn the mapping relationship from historical data to future data. In this way, the model can be trained to predict the amplitude of body surface fluctuations at a specific future time point, rather than simply identifying the current breathing stage. The choice of the fixed time step should be optimized based on the required prediction lead and control cycle in the actual application to ensure that the prediction results can be used in a timely manner for compensatory control.

[0036] Reference Figure 3 As shown, in a preferred embodiment of the present invention, in order to accurately capture the subtle patterns of body surface fluctuations during a patient's breathing process, the present invention further defines the specific implementation method for extracting temporal phase features in the breathing pattern prediction model. It should be noted that traditional sliding window averaging or simple peak detection algorithms struggle to distinguish between large trunk displacements caused by deep breathing and sudden muscle tremors caused by coughing. This embodiment introduces a joint architecture of dilated convolution and gated linear units, enabling the model to "sense" the specific stage of the entire respiratory cycle at the current moment (e.g., early inhalation, late inhalation, or apnea). The specific extraction process is described in detail below: First, the preprocessed body surface displacement data sequence x(t) is input into the encoding layer of the temporal convolutional network. In the encoding layer of the aforementioned respiratory pattern prediction model, a dilated convolution operation is performed on the input body surface displacement data sequence x(t). The dilated convolution operation, by introducing a preset interval (controlled by the dilation factor d) between the convolution kernel elements, can effectively expand the receptive field without increasing computational complexity, thereby capturing a longer range of temporal dependencies in the body surface displacement data sequence. This is crucial for understanding the periodicity and multi-scale characteristics of respiratory motion. The output F(t) of the dilated convolution operation at time t satisfies the following relationship: ;

[0037] Where Wm is the weight parameter of the convolution kernel at position m, M is the kernel size, d is the dilation factor, and b is the bias term. The dilation factor d allows the convolution kernel to skip some input data, thus effectively acquiring contextual information over a longer time span without increasing the number of parameters. This is particularly advantageous for capturing long-term patterns in the respiratory cycle. In this embodiment, d is preferably set to 2 or d=4. When d=2, the above operation essentially reads data once every sampling point. The weight parameter Wm and the bias term b are determined through training with pre-collected respiratory sample data, and during training, they automatically learn the feature extraction capabilities sensitive to inspiratory and expiratory slopes.

[0038] Subsequently, the output F(t) of the dilated convolution operation is input into a gated linear unit for nonlinear activation to generate a high-dimensional feature vector H. The gated linear unit, by introducing a gating mechanism, can adaptively control the information flow, thereby filtering out more discriminative features from complex time-series data and suppressing noise interference. This high-dimensional feature vector H satisfies the following relationship: ;

[0039] Here, F1(t) and F2(t) are the results of dual-channel linear projection of F(t): F1(t) is the information path, carrying specific feature information; F2(t) is the control path, after being processed by the Sigmoid function σ, the output value is between (0,1), which plays the role of a soft switch or gating signal; ⊙ represents element-wise multiplication operation.

[0040] F1(t) contains all the potential fluctuation features extracted from the original signal. σF2(t) assigns a weight coefficient between (0,1) to each feature element in F1(t) based on the current signal context. For example, when the input signal exhibits a slow and uniform waveform (corresponding to a calm inhalation phase), the weights of low-frequency components in σF2(t) approach 1, while the weights of high-frequency jitter components approach 0. Through this gating mechanism, the model can dynamically determine which feature information needs to be retained and transmitted, and which needs to be suppressed, thus enabling the generated feature vector H to more accurately focus on the key phase information of the respiratory motion. Ultimately, the generated high-dimensional feature vector H represents the temporal phase characteristics of the current respiratory motion phase.

[0041] Reference Figure 4 As shown, based on the construction of a respiratory pattern prediction model using a temporal convolutional network with an encoder-decoder architecture, this application further proposes a step to infer and output the expected body surface fluctuation amplitude of the next control cycle based on the intrinsic correlation between temporal phase characteristics and historical fluctuation amplitude variation patterns. Specifically, the steps include: in the decoding layer, a weighted summation operation is performed on the high-dimensional feature vector using a fully connected layer. The weight parameters of the fully connected layer are pre-fitted with the slope and extreme values ​​of body surface displacement changes corresponding to different respiratory phase stages using a training sample set. Based on the result of the weighted summation operation, a scalar value representing the expected body surface fluctuation amplitude of the next control cycle is directly generated.

[0042] Specifically, the decoding layer is a key component of the breathing pattern prediction model (a temporal convolutional network employing an encoder-decoder architecture). Its main function is to receive the high-dimensional feature vector H output by the encoding layer and convert it into a prediction output with actual physical meaning. In this application, the decoding layer is responsible for mapping the temporal phase features representing the current stage of respiratory movement to the expected amplitude of body surface fluctuations in the next control cycle. The fully connected layer, as the core component of the decoding layer, is a neural network layer in which each input neuron is connected to each output neuron. It achieves feature combination and mapping by performing a linear transformation (i.e., weighted summation) on the input data. In this application, the fully connected layer is used to receive the high-dimensional feature vector H from the encoding layer and synthesize these features through its internal weight parameters to generate a single scalar value as the prediction result. The weighted summation operation on the high-dimensional feature vector is the core operation of the fully connected layer, which multiplies each element of the high-dimensional feature vector H with the corresponding weight parameter and sums all products. This process allows the model to learn and identify which parts of the high-dimensional feature vector are most important for predicting the expected amplitude of body surface fluctuations, and assign different weights according to their importance, thereby effectively extracting key information related to the target prediction value from complex features.

[0043] The weight parameters of the fully connected layer are not randomly set, but are obtained in advance through learning and optimization on a pre-built training sample set. This training sample set fits the slope and extreme values ​​of body surface displacement changes corresponding to different respiratory phases. This fitting process allows the weight parameters to capture the dynamic characteristics of body surface displacement changes under different respiratory phases (such as inspiration, expiration, and pause), including the rate of change (slope) and the maximum / minimum displacement reached (extreme values), thus enabling the model to more accurately understand and predict the trend and amplitude of respiratory movements. Finally, based on the result of the weighted summation operation, the fully connected layer directly outputs a single scalar value, which is the system's prediction of the amplitude of body surface fluctuations in the next control cycle, ensuring the simplicity and operability of the prediction results.

[0044] It is important to note that the extraction of temporal phase features and amplitude inference processes described above simultaneously endow the breathing pattern prediction model with generalization ability across different individual body types. This generalization ability stems from the following two key technical features in the model architecture design: First, the temporal phase features extracted by the encoding layer through dilated convolution operations are an abstract representation of the normalized rhythmic pattern of respiratory movements. Specifically, the control pathway in the gated linear unit, after being processed by the Sigmoid function, dynamically generates gate weight coefficients within the (0,1) interval based on the temporal context of the input signal. When the input signal exhibits a slow and uniform waveform, the weights corresponding to low-frequency components approach 1, and the weights corresponding to high-frequency jitter components approach 0. This mechanism allows the model to focus on the phase progression pattern of respiratory movements—that is, the relative temporal relationship between the initial and final stages of inspiration, the initial stage of expiration, and the final stage of expiration—rather than the specific numerical values ​​of absolute displacement amplitude. Since different subjects, even if the absolute amplitude of body surface undulations varies due to differences in body size, still possess cross-individual physiological commonalities in the phase progression pattern of respiratory movements, this feature extraction method ensures the model's basic adaptability to subjects of different body types from an architectural perspective.

[0045] It should be further clarified that, to ensure the model's generalization ability across individuals from the outset, the construction of the training sample set described in this application follows the principle of feature coverage rather than sample exhaustion. Specifically, while collecting data on various modes such as calm breathing, deep breathing, and coughing, it is necessary to ensure that the data has sufficient variance across multiple quantifiable respiratory physiological feature dimensions, such as respiratory rate, inspiratory-to-expiratory ratio, and chest-abdominal displacement contribution ratio. By incorporating differentiated respiratory waveforms from subjects of different body types, ages, and genders into the training, the encoding layer of the temporal convolutional network can autonomously learn high-dimensional feature vectors that characterize the intrinsic rhythm of respiratory movements—i.e., temporal phase features—that are insensitive to the aforementioned individual morphological differences. Therefore, the model learns the dynamic process of breathing, rather than the thoracic morphology function of a specific individual.

[0046] Secondly, the fully connected layers in the decoding layer establish a mapping relationship from normalized phase features to individualized displacement amplitudes by pre-fitting the slope and extreme values ​​of body surface displacement changes corresponding to different respiratory phase stages on a training sample set covering various respiratory states. When subsequent online calibration is introduced, only the bias term needs to be fine-tuned to complete the offset correction of the mapping relationship, without retraining the entire network. This layered strategy of "shared feature extraction and adapted output mapping" enables the model to quickly adapt to different subjects with extremely low computational cost.

[0047] Specifically, to address the model prediction bias caused by differences in individual body shape and breathing habits among different examinees, the adaptive control method for ultrasound probe contact pressure based on respiratory compensation includes an online calibration phase before the formal examination procedure, comprising the following steps: While the subject is in a calm state, a series of body surface fluctuation displacement data are continuously collected for a preset duration as calibration data. The preset duration is preferably 30 to 60 seconds to ensure that at least 3 to 5 complete respiratory cycles can be captured. The calibration data is input into a pre-constructed respiratory pattern prediction model. The encoding layer extracts the individualized respiratory phase features corresponding to the calibration data, and the decoding layer outputs the corresponding predicted amplitude sequence. With the goal of minimizing the mean square error between the predicted amplitude of the decoding layer and the actual acquired amplitude, the bias term parameters of the fully connected layer in the decoding layer are fine-tuned and updated, while keeping the weight parameters of the dilated convolution kernel in the encoding layer unchanged. The above iterative process is repeated until the prediction error converges to a preset range or reaches a preset maximum number of iterations, thereby completing the personalized adaptation of the respiratory pattern prediction model to the current subject. The fine-tuned model parameters are then applied to the inference output step in the subsequent formal examination.

[0048] It is important to note that the fine-tuning of the bias term parameters keeps the weights of the encoding layer unchanged, adjusting only the bias term of the decoding layer. This design ensures that the respiratory phase features extracted by the model—i.e., the normalized rhythmic pattern output by the encoding layer—remain shared across different individuals, while the mapping between output amplitude and individual absolute displacement is adjusted only through the adaptation of the bias term. This strategy of shared feature extraction and adapted output mapping ensures that the model fully utilizes common respiratory patterns while achieving precise adaptation to the amplitude of body surface fluctuations for different individuals.

[0049] The working principle of the aforementioned online calibration phase can be explained from a technical perspective as follows: Human respiratory movements exhibit significant commonalities and characteristics at the individual level. The commonalities lie in the fact that, regardless of the subject's body type, age, or breathing habits, a single respiratory cycle can be divided into four basic phases: early inhalation, late inhalation, early exhalation, and late exhalation. The relative temporal proportions between these phases show statistical regularity. The characteristics lie in the fact that, due to individual factors such as thoracic volume, preference for abdominal / thoracic breathing, and subcutaneous fat thickness, different subjects exhibit significant differences in the absolute displacement of the body surface during the same respiratory phase.

[0050] Based on the distinction between the aforementioned commonalities and unique characteristics, this application adopts a layered strategy of "extracting common phase features at the coding layer and adapting characteristic amplitude mapping at the decoding layer." During the online calibration phase, only the bias term of the decoding layer is fine-tuned, which is equivalent to adjusting only the scale factor and offset of the output amplitude without changing the model's ability to recognize respiratory phases. This process is similar to zero-point calibration and gain calibration performed on different measurement channels in traditional signal processing; its essence is to complete the offset correction of the linear mapping layer with an extremely small sample size.

[0051] Based on the above description, those skilled in the art can complete the adaptation simply by using a conventional deep learning framework, setting the decoding layer bias term to a trainable state and the encoding layer weights to a frozen state during the online calibration phase, and performing several iterations of training using the mean squared error loss function. The entire calibration process can be completed within seconds on a conventional embedded processor, without significantly increasing the inspection preparation time. After calibration, the system enters the real-time adaptive control phase. In this phase, the breathing pattern prediction model performs phase feature extraction and amplitude inference based on the fine-tuned parameters, and uses the inference results for subsequent compensation calculation and control command generation.

[0052] Traditional compensation control methods typically use a fixed proportional coefficient for adjustment, which ignores the continuous change of respiratory motion over time. In reality, the fluctuations on the patient's body surface are not abrupt changes, but a dynamic process that changes continuously over time. If compensation is based solely on a single amplitude value at the current moment, the robotic arm's movement will lag behind the changes in the body surface, resulting in a catch-up phenomenon and exacerbating fluctuations in contact pressure. To address this issue, this embodiment uses a first-order linear differential function to generate the displacement change trend function based on the time phase characteristics of the respiratory pattern prediction model. This transforms the discrete predicted values ​​into a continuous motion trajectory description, thereby achieving proactive compensation for body surface fluctuations. Specifically, the target position compensation amount ΔL of the probe used to counteract the effects of breathing satisfies the following relationship: ; Where t0 is the current time, T is the duration of the next control cycle, and Apred is the expected amplitude of body surface fluctuations. dt is the rate of change of respiratory phase determined based on time phase characteristics, α is a preset displacement compensation proportional factor, and dt is the integral variable.

[0053] After obtaining the high-dimensional feature vector H of the time phase characteristics of the respiratory pattern prediction model output and the expected body surface fluctuation amplitude Apred, the system first constructs a displacement change trend function to describe the displacement change rate and direction of the body surface at multiple consecutive moments in the future.

[0054] The displacement change trend function is a first-order linear differential function, and its output is a gradient vector describing the rate of change of body surface displacement. This function is constructed based on the following physiological prior knowledge: during respiration, the rate of change of body surface displacement is closely related to the current respiratory phase. For example, the rate of change of displacement is relatively large at the beginning of inspiration and expiration; while at the end of inspiration and expiration, the rate of change of displacement approaches zero.

[0055] Based on the respiratory phase encoding information contained in the high-dimensional feature vector H, the system can extract the respiratory phase change rate at the current moment. Respiratory phase change rate The physical meaning of is: the rate at which the respiratory cycle phase advances per unit time, and its magnitude reflects the degree of rapid breathing in the patient.

[0056] Specifically, respiratory phase change rate The high-dimensional feature vector H can be obtained by inputting it into a fully connected regression layer, which outputs a scalar value representing the rate of change of the current phase relative to the previous moment. The weight parameters of this fully connected regression layer are determined during the model training phase by fitting phase change samples at different breathing frequencies.

[0057] The integrand describes the rate of change of the body surface displacement at any infinitesimal moment. By integrating this rate over the control cycle duration T, the cumulative displacement change of the body surface in the next control cycle can be obtained. Since the probe needs to move synchronously with the body surface to maintain contact stability, the magnitude of the probe's target position compensation is equal to this cumulative displacement change, and its direction is consistent with the direction of the body surface movement.

[0058] It is important to note that the displacement compensation scaling factor α is set considering the mechanical coupling characteristics between the probe and the body surface. When α = 1.0, it means the probe completely follows the movement of the body surface; when α < 1.0, it means the probe's following amplitude is slightly smaller than the body surface's movement amplitude, suitable for scenarios with thicker soft tissue and buffer space; when α > 1.0, it means the probe applies a slight pre-pressure to follow, suitable for scenarios with prominent bones on the body surface where sufficient contact needs to be ensured.

[0059] By defining the displacement change trend function as a first-order linear differential function and calculating the target position compensation amount ΔL using integral form, this application can more accurately predict and compensate for the dynamic displacement of the subject's surface during respiratory movements. Specifically, based on the known expected surface undulation amplitude and respiratory phase change rate, this method can calculate the precise displacement compensation amount required by the probe in the next control cycle in real time and dynamically by integrating the respiratory phase change rate. This dynamic trend-based compensation method overcomes the lag or insufficiency problems that may occur when compensating based solely on static amplitude prediction. This allows the robotic arm to drive the probe more smoothly and accurately to follow the surface undulations, thereby maintaining continuous and stable contact with the subject's surface while effectively avoiding contact pressure fluctuations caused by respiratory movements, significantly improving the stability and diagnostic quality of ultrasound imaging.

[0060] Furthermore, after obtaining the target position compensation amount, this application proposes a method for calculating the target pressure adjustment amount ΔF to achieve precise adaptive control of the contact pressure. Specifically, the target pressure adjustment amount ΔF satisfies the following relationship: ;

[0061] Where β is the preset force-position coupling coefficient; The mean value of the contact pressure distribution data at the current moment is obtained by calculating the arithmetic mean of the pressure values ​​of all valid sampling points in the sensor array.

[0062] Among these parameters, the preset force-position coupling coefficient β is a key parameter, characterizing the intrinsic relationship between the force applied by the probe and the displacement or deformation of the body surface. The value of the force-position coupling coefficient β can be adaptively adjusted according to the elastic modulus of the soft tissue in different examination sites. For example, for areas rich in soft tissue such as the abdomen, a larger β value can be selected to enhance pressure-following response; for superficial organ areas such as the thyroid gland, a smaller β value can be selected to ensure examination comfort. This coefficient usually needs to be obtained through prior experimental calibration. For example, under controlled conditions, a known force is applied to different subjects, and the resulting surface deformation or probe displacement is measured to establish a mathematical model between force and displacement. The accuracy of this coefficient directly affects the precision of pressure regulation; it can convert the expected displacement change trend caused by respiratory movements into the required mechanical compensation.

[0063] Mean of the pressure distribution data at the current moment The pressure is obtained by real-time monitoring of the actual contact pressure between the probe and the surface of the body being examined. For example, a flexible piezoresistive thin-film sensor array integrated into the front end of the ultrasound probe can be used to collect pressure distribution data of the contact surface between the probe and the body surface in real time. By averaging these distribution data, the average contact pressure at the current moment can be obtained. This real-time feedback information is crucial for closed-loop pressure control, as it reflects the actual force condition of the probe in contact with the body surface and provides a benchmark for subsequent pressure adjustment.

[0064] During the contact between the ultrasound probe and human skin, they form an approximately linear spring-damped system. According to Hooke's Law, the change in contact pressure is proportional to the change in probe displacement. Therefore, when the probe needs to move a distance ΔL to follow the undulations of the body surface, if the applied pressure is not adjusted accordingly, the contact pressure will change proportionally to the displacement.

[0065] The sign of the target pressure adjustment amount ΔF is related to the sign of the target position compensation amount ΔL. Specifically: When ΔL is positive (indicating that the probe needs to be moved closer to the body surface), ΔF is positive, indicating that the pressure needs to be increased to compensate for the increase in natural pressure caused by the bulge of the body surface and to prevent the probe from excessively compressing the tissue. When ΔL is negative (indicating that the probe needs to be moved away from the body surface), ΔF is negative, indicating that the applied pressure needs to be reduced to compensate for the natural pressure reduction caused by the depression of the body surface, and to prevent the probe from detaching from the skin.

[0066] Through the above technical solution, this application enables precise adaptive control of the contact pressure of the ultrasound probe. While calculating the target position compensation amount ΔL to counteract the effects of breathing, the system further dynamically calculates the target pressure adjustment amount based on the expected surface fluctuation amplitude and respiratory phase change rate output by the respiratory pattern prediction model, combined with a preset force-position coupling coefficient and the average value of the current contact pressure distribution data acquired in real time. This pressure adjustment mechanism based on respiratory motion prediction and real-time pressure feedback allows the robotic arm to actively adjust the force applied at its end while driving the probe for position compensation, effectively counteracting pressure fluctuations caused by respiratory motion. This not only ensures that the probe maintains stable contact with the patient's surface throughout the entire respiratory cycle, avoiding tissue damage or discomfort caused by excessive pressure, but also prevents poor contact or ultrasound signal attenuation caused by insufficient pressure, greatly improving the stability and diagnostic accuracy of ultrasound imaging. Simultaneously, by coordinating displacement compensation and pressure adjustment, comprehensive and precise control of probe movement and contact state is achieved, significantly improving the reliability of automated ultrasound examinations and the user experience.

[0067] In some embodiments described above, when the expected amplitude of body surface fluctuations exceeds a preset steady-state threshold, the system calculates a target position compensation amount and a target pressure adjustment amount to counteract the respiratory effects. However, in practical applications, the respiratory state of the subject may not always be stable. For example, when a patient experiences tension, pain, or a pathological condition, the regularity of their body surface fluctuations may change, leading to greater fluctuations in the contact pressure distribution data. If a fixed steady-state threshold is still used in this case, it may not be able to capture such changes in a timely and accurate manner, thereby affecting the sensitivity of the control system to respiratory movements, resulting in a decrease in the accuracy of probe-body surface contact pressure control, and even the possibility of the probe detaching from the body surface or applying excessive pressure.

[0068] In this regard, this application further proposes that the preset steady-state threshold is dynamically adjusted based on the pressure fluctuation variance of multiple sampling points in the contact pressure distribution data within a preset time window. If the pressure fluctuation variance increases, the value of the steady-state threshold is reduced so that the control method is more sensitive to surface fluctuations.

[0069] Specifically, the preset steady-state threshold is a key parameter used to determine whether the expected amplitude of body surface fluctuations needs to trigger a compensation mechanism. Initially, a baseline value can be set based on the routine requirements of ultrasound examinations or the physiological characteristics of specific areas. However, to accommodate the complexities of actual operation, this threshold is not fixed but has the ability to be dynamically adjusted.

[0070] Contact pressure distribution data is acquired in real time by a sensor array, reflecting the contact state between the probe and the subject's surface. This data includes multiple spatial sampling points, each generating a series of pressure readings within a consecutive preset time window. Pressure fluctuation variance is the result of statistical analysis of these time-series and spatial distribution data, quantifying the instability or dispersion of the contact pressure. For example, the overall pressure fluctuation variance can be obtained by calculating the variance of the pressure values ​​at each sampling point within a preset time window, and then averaging or weighted averaging the variances of all sampling points. When the subject's breathing is irregular or other disturbances occur, the contact pressure distribution will exhibit greater fluctuations, leading to an increase in pressure fluctuation variance.

[0071] The dynamic adjustment mechanism means that the system can optimize its control strategy based on real-time physiological feedback (i.e., pressure fluctuation variance). When an increase in pressure fluctuation variance is detected, it indicates a decrease in the stability of the current probe contact with the body surface, which may foreshadow more intense or irregular respiratory movements in the subject. At this time, the system will immediately lower the preset steady-state threshold. This strategy of lowering the steady-state threshold means that even if the expected fluctuations in the body surface are relatively small, as long as they exceed the lowered threshold, the system will immediately activate the compensation mechanism, making the control method more sensitive to body surface fluctuations.

[0072] In some of the above embodiments, although the target position compensation and target pressure adjustment of the probe are calculated based on respiratory motion, how to effectively transform these abstract compensation values ​​into low-level control commands that the robotic arm actuator can understand and execute precisely, so as to ensure that the probe can accurately track the undulations of the body surface and stably maintain contact pressure during complex movements, remains a technical problem that needs to be solved. Without a sophisticated command generation mechanism, the robotic arm may struggle to achieve smooth and precise follow-up and force control, affecting the stability and safety of ultrasound examinations.

[0073] Reference Figure 5 As shown, this application further proposes steps for generating a sequence of control commands to drive the actuators of a robotic arm, specifically including the following aspects: First, inverse kinematics calculation is performed on the target position compensation amount to obtain the target rotation angles of each joint of the robotic arm. Inverse kinematics calculation is a key technology in robot control. Its role is to calculate the specific angles or displacements that each joint of the robotic arm needs to achieve based on the desired position and orientation of the end effector (i.e., the ultrasonic probe) in Cartesian space. For example, numerical iterative methods, such as the Jacobian matrix iteration method, can be used to solve complex nonlinear kinematic equations by continuously approximating them, thereby accurately converting the displacement compensation required by the probe (e.g., moving X millimeters in a certain direction) into specific rotation commands for each joint of the robotic arm (e.g., shoulder joint, elbow joint, wrist joint). This ensures that the robotic arm can accurately move the probe to the expected compensation position to counteract the effects of surface undulations.

[0074] Secondly, the target pressure adjustment is converted into impedance control parameters at the robotic arm's end effector. The target pressure adjustment is the force that needs to be applied or adjusted to maintain the contact pressure between the probe and the body surface within a preset range. Impedance control is an advanced force / position hybrid control strategy that enables the robotic arm's end effector to exhibit the expected mechanical impedance characteristics when in contact with the environment; that is, the dynamic relationship between force and displacement conforms to a preset stiffness, damping, and inertia model. By mapping the target pressure adjustment to impedance control parameters, such as dynamically adjusting the virtual stiffness or target force at the robotic arm's end effector, the robotic arm can smoothly interact with the patient's body surface. For example, when increased contact pressure is needed, the virtual stiffness can be appropriately increased or the target force adjusted, and vice versa. This effectively manages the contact force while maintaining contact, avoiding excessive or insufficient pressure that could cause discomfort to the patient or affect the quality of ultrasound images.

[0075] Furthermore, the generated control command sequence contains interpolated trajectory data of the target rotation angle and impedance control parameters. The control command sequence is the underlying command executed by the robotic arm controller; it includes not only the target rotation angles that each joint of the robotic arm needs to achieve, but also the impedance control parameters that need to be dynamically adjusted throughout the movement. The interpolated trajectory data refers to a series of smooth intermediate points generated between consecutive target points using mathematical methods (such as spline interpolation or polynomial interpolation). This means that when the robotic arm moves from the current state to the next compensated state, its joint angles and impedance control parameters will gradually change according to a preset smooth curve, rather than undergoing abrupt changes. This interpolation process effectively avoids potential shocks and vibrations during robotic arm movement, significantly improving the smoothness and accuracy of the movement, while ensuring that the contact pressure between the probe and the body surface is effectively and smoothly controlled throughout the entire movement, avoiding pressure fluctuations caused by sudden changes in commands.

[0076] Through the above technical solution, this application can efficiently and accurately convert abstract target position compensation and target pressure adjustment into low-level control commands executable by the robotic arm. Inverse kinematics calculation ensures that the probe can accurately track the respiratory fluctuations of the examined body surface, while the introduction of impedance control parameters makes the robotic arm exhibit compliant mechanical characteristics when in contact with the body surface, effectively avoiding excessive or insufficient pressure caused by rigid control, thus maintaining stable contact pressure during dynamic contact. In addition, the interpolated trajectory data included in the control command sequence further ensures the smoothness and continuity of the robotic arm's movement, avoiding motion shocks and pressure abrupt changes, and significantly improving the accuracy and stability of probe follow-up and force control. This refined command generation mechanism enables the robotic arm to perform respiratory compensation tasks efficiently and safely, thereby significantly improving the stability and comfort of probe contact with the body surface during ultrasound examination, and thus improving the quality of ultrasound images and the accuracy of diagnosis.

[0077] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for adaptive control of ultrasonic probe contact pressure based on respiratory compensation, characterized in that, Includes the following steps: The sensor array acquires the surface undulation signal of the tested body and the probe contact pressure signal. After filtering and preprocessing, it generates real-time surface undulation displacement data sequence and contact pressure distribution data. The body surface fluctuation displacement data sequence is input into a pre-constructed respiratory pattern prediction model. The respiratory pattern prediction model is constructed using a temporal convolutional network with an encoder-decoder architecture, including an encoding layer and a decoding layer. The respiratory pattern prediction model extracts the temporal phase features representing the current respiratory movement stage, including: in the encoding layer, performing a dilated convolution operation on the input body surface fluctuation displacement data sequence, and inputting the output of the dilated convolution operation into a gated linear unit for nonlinear activation to generate a high-dimensional feature vector representing the temporal phase features of the current respiratory movement stage; based on the inherent correlation between the temporal phase features and the historical fluctuation amplitude change pattern, the expected body surface fluctuation amplitude for the next control cycle is inferred and output, including: in the decoding layer, using a fully connected layer to perform a weighted summation operation on the high-dimensional feature vector. The weight parameters of the fully connected layer are pre-fitted with the slope and extreme values ​​of the body surface displacement change corresponding to different respiratory phase stages through a training sample set. Based on the result of the weighted summation operation, a scalar value representing the expected body surface fluctuation amplitude for the next control cycle is directly generated. In response to the expected fluctuation amplitude of the body surface exceeding the preset steady-state threshold, the displacement change trend function generated by the breathing pattern prediction model based on the time phase characteristics is used to calculate the target position compensation amount of the probe used to counteract the breathing effect, and the target pressure regulation amount is calculated based on the target position compensation amount. Wherein: the displacement change trend function is a first-order linear differential function, and the target position compensation amount ΔL of the probe used to offset the effect of breathing satisfies the following relationship: ; Where t0 is the current time, T is the duration of the next control cycle, and Apred is the expected amplitude of body surface fluctuations. The respiratory phase change rate is determined based on the time phase characteristics, α is a preset displacement compensation scaling factor, and dt is the integral variable; The target pressure adjustment ΔF satisfies the following relationship: ; Where β is the preset force-position coupling coefficient; This represents the mean of the contact pressure distribution data at the current moment. Based on the target position compensation amount and the target pressure adjustment amount, combined with the probe contact pressure distribution data, a sequence of control commands is generated to drive the robotic arm actuator, so that the robotic arm drives the probe to maintain contact with the surface of the object being examined while keeping the contact pressure within a preset range.

2. The adaptive control method for ultrasonic probe contact pressure based on respiratory compensation according to claim 1, characterized in that: The sensor array includes a flexible piezoresistive thin-film sensor and an inertial measurement unit integrated into the front end of the ultrasonic probe. The displacement data sequence of body surface undulations is obtained by calculating the acceleration and angular velocity data output by the inertial measurement unit, and the contact pressure distribution data is obtained by analyzing the resistance change array of the flexible piezoresistive thin-film sensor.

3. The adaptive control method for ultrasonic probe contact pressure based on respiratory compensation according to claim 1, characterized in that: The encoding layer is used to extract deep features from the body surface undulation displacement data sequence, and the decoding layer is used to map the deep features into displacement prediction values ​​for future times.

4. The adaptive control method for ultrasonic probe contact pressure based on respiratory compensation according to claim 3, characterized in that: The training sample set used in the pre-construction phase of the breathing pattern prediction model is constructed by continuously collecting body surface displacement data sequences when the patient is in a state of calm breathing, deep breathing, and coughing. The label data of the training sample set is the actual body surface displacement value corresponding to a fixed time step after the collection time.

5. The adaptive control method for ultrasonic probe contact pressure based on respiratory compensation according to claim 3, characterized in that: The output F(t) of the dilated convolution operation at time t satisfies the following relationship: ； Where: Wm is the weight parameter of the convolution kernel at the m position, M is the convolution kernel size, d is the dilation factor, and b is the bias term; The output of the dilated convolution operation is input into a gated linear unit for nonlinear activation, generating a high-dimensional feature vector H that satisfies the following relationship: ； Here, F1(t) and F2(t) are the results of performing a two-channel linear projection on F(t): F1(t) is the information path, carrying specific feature information; F2(t) is the control path, and after being processed by the Sigmoid function σ, the output value is between (0,1); ⊙ represents the element-wise multiplication operation.

6. The adaptive control method for ultrasonic probe contact pressure based on respiratory compensation according to claim 1, characterized in that: The preset steady-state threshold is dynamically adjusted based on the pressure fluctuation variance of multiple sampling points in the contact pressure distribution data within a preset time window. If the pressure fluctuation variance increases, the value of the steady-state threshold is reduced so that the control method is more sensitive to surface fluctuations.

7. The adaptive control method for ultrasonic probe contact pressure based on respiratory compensation according to claim 1, characterized in that: The steps for generating a control command sequence to drive the robotic arm actuator specifically include: performing inverse kinematics calculation on the target position compensation amount to obtain the target rotation angle of each joint of the robotic arm, and converting the target pressure adjustment amount into impedance control parameters at the end of the robotic arm. The control command sequence includes interpolated trajectory data of the target rotation angle and the impedance control parameters.

Images