Ice blanket machine dynamic temperature control method based on body temperature feedback

Abstract

The application relates to the technical field of medical temperature control, and discloses an ice blanket machine dynamic temperature control method based on body temperature feedback. The method acquires body surface temperature sensing data and core temperature simulation data of a target object through multi-source sensing of body temperature, and constructs an original body temperature sensing data set; then physiological state feature vectors are generated according to the data for physiological state evaluation; based on the feature vectors, a multi-modal body temperature state tensor is generated through multi-source sensing fusion modeling; subsequently, a reinforcement learning method dynamically bound with a clinical constraint rule component is adopted to generate an optimized temperature control strategy according to the multi-modal body temperature state tensor; the optimized strategy is subjected to integrity verification and clinical safety rule conflict detection to generate a safe control instruction sequence; finally, the sequence is used to drive the ice blanket machine to perform temperature regulation operation. The method improves the accuracy and safety of ice blanket machine temperature regulation and adapts to complex clinical scene requirements.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical temperature control, in particular to an ice blanket machine dynamic temperature control method based on body temperature feedback. BACKGROUND

[0002] In clinical treatment, ice blanket machines, as an important body temperature regulation equipment, are widely used in scenarios such as cooling of hyperthermia patients and postoperative body temperature maintenance. Its core function is to intervene in the patient's body temperature by regulating the blanket surface temperature, so as to achieve the purpose of treatment or auxiliary treatment. However, the current temperature control method of ice blanket machine still has many limitations, which is difficult to meet the precise regulation demand in complex clinical scenarios.

[0003] The temperature sensing method of the existing ice blanket machine mainly depends on a single body surface temperature sensor, which can only obtain the temperature data of the patient's local body surface. However, the body temperature distribution of human body has significant complexity, and there is a large difference between the body surface temperature and the core temperature, especially when the patient is in a special physiological state such as fever, shock or anesthesia. The single body surface temperature data cannot fully reflect the real body temperature state of the patient, which may lead to the deviation of the temperature control decision from the actual demand, for example, the cooling intensity may be reduced too early due to the temporary decrease of the body surface temperature, while the core temperature is still at a high level.

[0004] In terms of temperature regulation strategy, traditional ice blanket machines mostly use preset fixed regulation mode, such as stepwise adjustment according to the set temperature threshold. This mode lacks the ability to dynamically adapt to the physiological state of individual patients, and there are significant differences in the basal metabolic rate, severity of illness, sensitivity to temperature, etc. among different patients, so it is difficult to take into account the individuality. At the same time, the existing strategy formulation process rarely systematically integrates clinical diagnosis and treatment rules, for example, for patients with bleeding tendency, excessive cooling may increase the risk of coagulopathy, but the existing regulation logic often does not consider such clinical constraints in decision-making, which may pose potential medical safety risks.

[0005] The control process of the existing ice blanket machine lacks a perfect safety verification mechanism. After the regulation strategy is generated, it directly drives the actuator to operate, without real-time detection of the integrity of the strategy and its compatibility with clinical safety rules. In complex clinical environments, factors such as abnormal sensor data, temporary equipment failure or sudden physiological state changes of patients may cause conflicts or omissions in the control instructions, thereby affecting the temperature control effect and even causing adverse events. For example, when the sensor data drifts, the control instructions generated based on the wrong data may make the blanket surface temperature too low, causing the patient to shiver and other discomfort reactions. SUMMARY

[0006] The present application aims to provide an ice blanket machine dynamic temperature control method based on body temperature feedback to solve the problems raised in the background.

[0007] To achieve the above object, the application provides a dynamic temperature control method of ice blanket machine based on body temperature feedback, which comprises the following steps:

[0008] Body temperature multi-source perception, obtaining the body surface temperature perception data and core temperature simulation data of the target object, and constructing the original body temperature perception data set;

[0009] Physiological state evaluation, generating a physiological state feature vector according to the body surface temperature perception data and the core temperature simulation data;

[0010] Multi-modal body temperature state modeling, based on the physiological state feature vector, generating a multi-modal body temperature state tensor through multi-source perception fusion modeling;

[0011] Body temperature regulation strategy optimization, generating an optimized temperature control strategy by using a reinforcement learning method dynamically bound with a clinical constraint rule component according to the multi-modal body temperature state tensor;

[0012] Control chain path topology verification, performing integrity verification and clinical safety rule conflict detection on the optimized temperature control strategy to generate a safe control instruction sequence;

[0013] Dynamic temperature execution, driving the ice blanket machine to perform temperature regulation operation according to the safe control instruction sequence.

[0014] Preferably, the body temperature multi-source perception comprises the following operations:

[0015] Collecting skin surface temperature time series through distributed temperature sensors and synchronously acquiring environmental temperature and humidity compensation parameters;

[0016] Generating core temperature simulation values through a core temperature estimation algorithm, combining the skin surface temperature time series and the environmental temperature and humidity compensation parameters, and constructing the original body temperature perception data set.

[0017] Preferably, the physiological state evaluation comprises the following operations:

[0018] Performing temperature time series correction on the original body temperature perception data set to generate a body surface temperature time series;

[0019] Inputting the body surface temperature time series into a physiological state mapping component to generate core temperature offset features through a body surface-core temperature correlation model;

[0020] Fusing the core temperature offset features and the environmental temperature and humidity compensation parameters to generate a physiological state feature vector.

[0021] Preferably, the multi-modal body temperature state modeling comprises the following operations:

[0022] A hierarchical attention fusion mechanism is constructed, including a body temperature adjustment response attention layer, an environmental interference suppression attention layer, and a clinical risk warning attention layer.

[0023] Temperature regulation demand features in the physiological state feature vector are extracted through the body temperature adjustment response attention layer.

[0024] Environmental temperature and humidity interference noise features are filtered through the environmental interference suppression attention layer.

[0025] A body temperature sudden change risk probability distribution is generated through the clinical risk warning attention layer.

[0026] The temperature regulation demand features, environmental temperature and humidity interference noise features, and body temperature sudden change risk probability distribution are aggregated to generate a multi-modal body temperature state tensor.

[0027] Preferably, the body temperature adjustment strategy optimization includes the following operations:

[0028] A reinforcement learning state parameter space is constructed, including an ice blanket current temperature state, a body temperature change rate state, and a clinical constraint rule component state.

[0029] A multi-layer policy network architecture is adopted, including an annual scheme generation layer, a clinical dynamic optimization layer, and a real-time execution decision layer.

[0030] A long-term body temperature management framework protocol is output through the annual scheme generation layer.

[0031] The long-term body temperature management framework protocol is dynamically adjusted through the clinical dynamic optimization layer to generate an hourly temperature control scheme.

[0032] Minute-level temperature adjustment instructions are generated through the real-time execution decision layer.

[0033] Preferably, the training of the multi-layer policy network architecture includes the following operations:

[0034] A multi-objective reward aggregation function is designed, integrating a temperature offset penalty factor, a thermal discomfort evaluation factor, and a device energy consumption factor.

[0035] A clinical disturbance simulation function is introduced to simulate body temperature abnormal fluctuation events, sensor failure events, and environmental mutation events.

[0036] The optimized temperature control strategy is generated through a phased optimization process, including an initial protocol generation phase, a clinical disturbance adaptation phase, and a multi-objective strategy aggregation phase.

[0037] Preferably, the control chain path topology verification includes the following operations:

[0038] constructing a control instruction dependency graph to record the execution dependency relationship of each temperature adjustment instruction in the optimized temperature control strategy;

[0039] detecting conflict nodes of the temperature adjustment instruction and the clinical safety rule component through a reverse reference mapping table;

[0040] checking the loop risk of the control instruction dependency graph by using a depth-first search algorithm;

[0041] generating a safe control instruction sequence without conflict and loop.

[0042] Preferably, the conflict detection of the clinical safety rule component comprises the following operations:

[0043] defining a body temperature safety boundary rule, a temperature change rate threshold rule and a device maximum power limit rule;

[0044] traversing each temperature adjustment instruction in the optimized temperature control strategy, and matching the violation identifier of the body temperature safety boundary rule, the temperature change rate threshold rule and the device maximum power limit rule through a bidirectional mapping tool;

[0045] deleting the temperature adjustment instruction containing the violation identifier.

[0046] Preferably, the generation of the safe control instruction sequence comprises the following operations:

[0047] sequencing the conflict-free temperature adjustment instruction according to the execution dependency relationship based on a topological sorting algorithm;

[0048] adding a device start-up preheating instruction and a sensor calibration instruction as a front node;

[0049] outputting the safe control instruction sequence arranged in sequence.

[0050] Preferably, the dynamic temperature execution comprises the following operations:

[0051] analyzing the temperature target value and the execution time window in the safe control instruction sequence;

[0052] driving an ice blanket machine compressor power regulation component and a circulating water flow control valve to gradually approach the temperature target value according to the execution time window;

[0053] real-time feedback of the deviation of the actual temperature value from the temperature target value to the body temperature multi-source perception step.

[0054] Compared with the prior art, the present application has the following beneficial effects:

[0055] In terms of body temperature perception, the method innovatively adopts a multi-source perception approach, simultaneously acquiring body surface temperature perception data and core temperature simulation data and constructing an original data set. This dual-dimensional perception mode breaks through the limitations of traditional single body surface temperature perception, enabling a more comprehensive capture of patient body temperature information. Body surface temperature directly reflects the instantaneous temperature state of the blanket action area, while core temperature simulation data reflects the deep body temperature of the body. The combination of the two can avoid misjudgment caused by single data deviation, enabling subsequent temperature control decisions to be based on more complete body temperature information.

[0056] The physiological state assessment step generates a feature vector through in-depth analysis of body surface and core temperature data, providing individualized basis for precise temperature control. Different patients have differences in age, disease, and physical fitness, and their physiological states also react differently to temperature stimulation. This step extracts features related to physiological state, converts abstract body temperature data into quantifiable state indicators, and enables the temperature control system to identify individual patient-specific characteristics, such as distinguishing between inflammatory response stages in high fever patients and metabolic recovery states in postoperative patients, and provides direction for subsequent strategy optimization.

[0057] Multi-modal body temperature state modeling generates a tensor by fusing multi-source perception data, further improving the accuracy of body temperature state representation. Body temperature changes are a dynamic process involving multiple factors interacting with each other, and single-dimensional features are difficult to fully describe the complex rules. Multi-modal modeling integrates body surface temperature dynamics, core temperature trends, and physiological feature changes into a unified tensor structure, which can more comprehensively depict the spatiotemporal distribution characteristics and evolution rules of patient body temperature state, providing more abundant state basis for the generation of regulation strategies, making the strategy more suitable for the actual body temperature change needs of patients.

[0058] The body temperature regulation strategy optimization step uses a reinforcement learning method with dynamic binding of clinical constraint rule components, making the regulation strategy more in line with clinical practice requirements. Reinforcement learning has the ability to optimize decision-making through dynamic learning, and the introduction of clinical constraint rules sets a safety boundary for the learning process. During strategy generation, the system can refer to clinical diagnosis and treatment specifications in real time, such as different disease temperature control targets and temperature tolerance ranges for special populations, so that the optimized strategy not only has dynamic adaptability but also complies with safety standards in clinical practice, avoiding regulation behaviors that do not conform to medical routines.

[0059] The control chain path topology verification adds a safety line to the temperature control operation. This link performs integrity check and safety rule conflict detection on the generated optimization strategy, which can timely discover possible omissions or contents that conflict with clinical safety requirements in the strategy. For example, when detecting a temperature sudden change instruction in the strategy, the safety of the instruction can be evaluated in combination with the cardiovascular state of the patient to avoid adverse reactions caused by improper instructions. By generating a safety control instruction sequence, the reliability of the regulation operation at the execution level is ensured, and the risk in clinical application is reduced.

[0060] The dynamic temperature execution link drives the device operation according to the safety instruction sequence, ensuring the immediacy and accuracy of the temperature control operation. After obtaining the safety instruction, the system can quickly drive the ice blanket machine to adjust the blanket surface temperature, so that the regulation measures can act on the patient in time. This fast response capability is particularly important for clinical scenarios with rapid body temperature fluctuations, and can timely follow the changes in the patient's body temperature, maintain the body temperature within the target range, and ensure the effectiveness of the temperature control effect. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 The timing diagram of the ice blanket machine dynamic temperature control method based on body temperature feedback according to the present application;

[0062] Figure 2 The physiological state assessment flowchart;

[0063] Figure 3 The multi-modal body temperature state modeling flowchart;

[0064] Figure 4 The control chain path topology verification flowchart. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0066] Please refer to Figure 1 The present application provides an ice blanket machine dynamic temperature control method based on body temperature feedback, which comprises:

[0067] The six core modules are multi-source temperature perception, physiological state assessment, multi-modal temperature state modeling, temperature regulation strategy optimization, control chain path topology verification, and dynamic temperature execution. The multi-source temperature perception module collects skin temperature data through distributed sensors, combines environmental parameters and core temperature estimation algorithms to construct the original data set. The physiological state assessment module generates a feature vector reflecting the physiological state through temperature time series correction and correlation model. The multi-modal temperature state modeling module uses a hierarchical attention mechanism to fuse temperature regulation requirements, environmental interference suppression, and clinical risk warning features. The temperature regulation strategy optimization module generates a hierarchical temperature control strategy based on the reinforcement learning framework. The control chain path topology verification module ensures the clinical safety of the instruction sequence through dependency analysis and conflict detection. The dynamic temperature execution module finally converts the safe instructions into specific regulation operations of the ice blanket machine, forming a closed-loop control system.

[0068] Example 1: refer to Figure 2 , covering two core processing stages of temperature data acquisition and physiological state conversion. The distributed temperature sensor array selects medical-grade platinum resistance elements and is deployed in six key anatomical regions of the target object's forehead, left and right carotid arteries, left and right axillary fossae, and sternum. The sensor nodes form a network through the Bluetooth 5.0 protocol, and use time division multiple access to achieve synchronous sampling with a sampling frequency of 10 times per second. Each sensor unit integrates a digital temperature and humidity composite chip, which captures air temperature and relative humidity parameters 2 cm away from the skin surface in real time while collecting skin temperature data. The humidity measurement range covers 20% to 95% RH. The data preprocessing unit automatically completes three-point calibration to eliminate sensor baseline drift errors and generates temperature raw data stream with time stamp.

[0069] The core temperature estimation algorithm uses a recurrent neural network as the calculation framework, constructing a three-layer long short-term memory network with 128 hidden layer units. The input layer receives four types of parameters: the temperature mean of the six regions, the maximum temperature gradient value between adjacent sensors, the heart rate variability index obtained from the medical monitor, and the environmental thermal resistance coefficient calculated according to the patient's covering thickness. The network hidden layer uses a gating mechanism to filter transient interference signals, and the output layer generates core temperature estimates and adds dynamic confidence interval calculation. The confidence interval is constructed based on the sliding window statistical method, with a window width of 300 consecutive sampling points. The probability distribution model is generated based on the standard deviation calculation. The original temperature perception data set uses a time-sharing storage structure, generating a data packet every 60 seconds. Each packet contains 26 fields such as the time stamp array accurate to milliseconds, the six sensor node numbers and their corresponding temperature-humidity values, the core temperature estimate and the upper and lower bounds of the confidence interval. The data storage format uses a lightweight binary encoding scheme, with each data packet fixed at 512 bytes in length.

[0070] In the physiological state assessment stage, the data correction engine starts the Kalman filter procedure. The process noise covariance matrix of the filter is dynamically adjusted according to the patient's activity state: when the accelerometer detects a change in body position, the noise tolerance value is automatically increased by 3 times; in the resting state, the basic covariance parameter is used. The temperature time series correction module establishes a dual processing channel: the main channel performs sliding window average processing on the original data, and the sub-channel uses median filtering to assist verification. A 30-second wide sliding window is advanced in 5-second steps, and the arithmetic mean of the data in the window is recalculated at each displacement. When the data difference between the main channel and the sub-channel exceeds 0.2 degrees Celsius, the abnormal data replacement mechanism is triggered.

[0071] The body-surface-core temperature association model is implemented using a Gaussian process regression framework. The input feature set includes five-dimensional parameters: the 120-second sliding average of the corrected body-surface temperature, the change slope of the body-surface temperature in the last 10 seconds, the confidence interval width of the core temperature estimate, the arithmetic mean of the ambient humidity reading, and the historical trend of the difference between the ambient temperature and the body-surface temperature. The model training stage loads the hyperparameter optimization module to determine the optimal combination of kernel functions through Bayesian tuning. The model outputs two key indicators: core temperature offset and its change rate, with a data update frequency of 2 times per second.

[0072] After the environmental temperature and humidity compensation parameters enter the feature fusion stage, the dynamic weighting algorithm is started. This algorithm automatically assigns weight coefficients according to the stability of the environmental sensor readings: when the environmental temperature fluctuation amplitude is less than 0.1°C for 3 consecutive samples, it is assigned a weight value of 0.3; when the minute-level change in humidity exceeds 5%, the weight value automatically decreases to 0.1. The fusion processor integrates three sources of data streams: core offset features from the temperature association model, environmental temperature and humidity weight values, and the third derivative features of the body-surface temperature itself. The feature vector constructor finally generates a 128-dimensional floating-point array, with each element storing a specific physiological state parameter in a fixed order. The first 64 dimensions carry temperature features, including body-surface temperature values, temperature difference matrices between different anatomical points, and temperature change acceleration parameters; the middle 32 dimensions store core temperature-related parameters, including offset, change trend confidence, and abnormal fluctuation count; the last 32 dimensions record environmental interaction parameters, including heat flux estimates, effective radiation area coefficients, and humidity influence factors. This feature vector is updated once every second and transmitted to the downstream processing unit through a double buffering mechanism to avoid processing delays caused by data transmission.

[0073] The data integrity monitoring system synchronously runs throughout the processing link. A numerical boundary checker is set at the temperature collection layer to discard abnormal data exceeding the physiological range of 34-42°C; a cyclic redundancy check technology is used at the transmission layer, with 16-bit check code added to each frame of data; and a null value detection unit is deployed at the feature vector generation layer to automatically enable a historical data interpolation mechanism when data is missing, with the interpolation algorithm selecting a linear extrapolation or mean filling strategy according to the data type difference. The entire processing flow adopts a time-triggered architecture, and all processing nodes are synchronized by a precise clock to ensure that the time alignment accuracy of data at each stage is controlled within 20 milliseconds.

[0074] Embodiment 2: see Figure 3 The body temperature state modeling module constructs a three-layer parallel attention processing architecture, in which the body temperature regulation response layer is configured with eight independent self-attention calculation units, each unit being equipped with a weight network of 512 neural nodes. These units use a key-value query mechanism to process different data subsets from the physiological state feature vector, focusing on the feature values reflecting the body surface heat flux in the 15th to 45th dimensions of the feature vector. The self-attention calculation is performed every 200 milliseconds, outputting a temperature regulation demand index matrix with a dimension of 8x64, each row representing the control demand intensity of a different body region.

[0075] The environmental interference suppression layer is implemented as a gated recurrent structure, composed of three layers of gated recurrent units with 128, 64, and 32 hidden units, respectively. This structure receives a total of 128-dimensional feature vector input while maintaining a length of 120 environmental noise template library in memory. The template matcher calculates the cosine similarity between the current environmental parameters and the historical noise templates in real time, dynamically loading the closest three templates to form an adaptive filter. The filtering parameters include temperature drift compensation coefficients and environmental decoupling factors, with an operation frequency of 10 Hz. The three groups of output gates connected through weight sharing mechanisms finally generate the reduced dimension anti-interference feature vector, with a dimension compressed to 16.

[0076] The clinical risk early warning layer is implemented as a probability generation network, with input ports receiving the body temperature regulation demand index matrix and the environmental filtered feature data. The risk predictor contains four parallel fully connected networks, with the output layer of each network using a Sigmoid activation function to calculate the probability estimates of low body temperature risk, high febrile convulsion risk, temperature drop risk, and device overload risk, respectively. The probability output range is limited to 0 to 1, with a resolution of 0.001, and each risk probability is updated independently with a response period of 0.5 seconds. The risk classifier is attached with a priority weighting mechanism, which automatically increases the weight of the device overload risk by 40% when multiple risks occur simultaneously.

[0077] The multi-modal tensor synthesizer receives the processing results of the three-layer attention mechanism: a 64-dimensional regulatory demand vector from the body temperature regulation layer, a 16-dimensional purification feature from the environmental inhibition layer, and a 4-dimensional probability vector from the risk warning layer. Through tensor product operation, the three types of data are projected into a unified high-dimensional space. The fusion processor adopts a cascade convolution operation, configures a 3x3 kernel and a step of 1, and applies three convolution transformations to output a 32-channel feature map. The tensor reconstruction unit finally generates a three-order tensor structure with dimensions of 16x16x16, and the tensor elements are stored in floating-point format with a dynamic range of -10.0 to +10.0.

[0078] In the adjustment strategy generation module, the state space of the reinforcement learning environment is accurately configured with 53 observation variables. The ice blanket temperature state adopts a five-level encoding mechanism: the 34-40°C interval is divided into 12 subintervals with a width of 0.5°C, and each subinterval has a state flag bit. The body temperature change rate parameter records the average change gradient in the last 30 seconds with a precision of 0.01°C / min. The clinical rule component state contains 19 binary indicators, which track the activation state of each safety constraint in real time.

[0079] The annual plan generation layer is implemented as a deep neural network with three input sources, including patient historical body temperature fluctuation patterns, environmental seasonal feature library, and clinical prescription constraints. The network structure contains seven fully connected layers with node numbers of 256, 128, 64, 32, 16, 8, and 8, respectively. The output layer generates 365 groups of temperature control reference values, each containing 96 preset temperature values for a single day, with a resolution of 0.1°C. The reference curve performs data smoothing every minute, using cubic spline interpolation to correct data jump points.

[0080] The clinical dynamic optimization layer uses a double-delayed policy gradient algorithm, with four independent value networks and two policy networks. The value network has 128 nodes per layer, and the policy network uses a 256-node hidden layer. The optimizer activates policy updates every 60 minutes, with inputs including real-time physiological state data, annual reference curve offset, and device working state logs. The optimizer output is a 240-element adjustment matrix, with each element corresponding to a temperature correction amount for the next 15 minutes, with a precision of 0.05°C.

[0081] The fine control branch of the decision layer constructs the policy gradient framework in real time, and realizes second-level response based on the proximal policy optimization method. The state encoder collects the multi-modal body temperature state tensor every 15 seconds, and compresses it into a 32-dimensional state vector through a feature extraction network. The execution policy network adopts a two-layer 256-node fully connected structure, and the output end is connected to three independent action channels: the temperature target value channel has a resolution of 0.01°C; the adjustment rate channel has a set value range of 0.1-5.0°C / min; and the expected stable time channel outputs an instruction value between 5-300 seconds in seconds. The action selector introduces a random exploration factor through Gaussian sampling, and the standard deviation is set to 5% of the target temperature difference.

[0082] Example 3: refer to Figure 4 , which covers the policy network training mechanism and the safety control instruction generation system. The reinforcement learning framework is configured with a multi-objective reward aggregation function :

[0083]

[0084] The function updates the parameter value at discrete time step t. represents the actual measured core temperature value, with a precision of 0.01°C; represents the target temperature value set by the clinic, which is allowed to float within the range of 36.5°C to 37.5°C; is the local thermal discomfort index calculated according to the ISO7730 standard, which is derived by combining the body surface temperature gradient and the sweat evaporation rate; is the normalized value of the instantaneous power consumption of the ice blanket compressor, with a reference of 2000W rated power. The weighting coefficient adopts an adaptive mechanism: automatically adjusts according to the body temperature offset amplitude, with a base value of 1.25 and an increase of 0.3 for every 0.5°C increase; maintains a constant value of 0.8; is dynamically adjusted within the range of 0.05-0.2 according to the device operating mode.

[0085] The network training is implemented in a phased optimization process. The initial protocol generation stage loads the historical case database, which contains 2000 complete body temperature regulation records of different physical conditions. The state transition model is constructed using the Markov decision process, and the transition probability matrix has a dimension of 53x53. The policy network is pre-trained for 15000 iterations, with 128 scenario data sampled each iteration. The optimizer uses the RMSprop algorithm with an initial learning rate of 3x10^{-5}. The basic policy parameter set is output in this stage and stored in the first 1024 sectors of the non-volatile memory.

[0086] The clinical disturbance adaptation stage constructs a virtual training environment. The body temperature abnormal fluctuation event simulator adopts a piecewise random walk model: set the maximum drift amplitude to ±0.2°C in the normal physiological state; when marked as pathological fluctuation, the drift amplitude is expanded to ±1.5°C. The sensor failure event injection module designs two failure modes: the random failure mode shields 10%-50% of the sensor nodes in a Bernoulli distribution; the regional failure mode specifies a group of sensors that continuously shield a specific anatomical area. The environmental mutation event generator applies a ±5°C step disturbance every 300 steps, with a random duration distributed between 30-180 seconds. The disturbance training executes a three-cycle mechanism: first, fine-tune the strategy parameters for 2000 iterations in a stable environment; then, inject three types of disturbances for 5000 iterations each; finally, perform combined disturbance training for 10000 iterations, with a total training step length of 3.2×10^5.

[0087] The multi-objective strategy aggregation stage starts the parameter fusion program. The value network constructs four parallel evaluation channels: the body temperature control channel calculates the average absolute error in a 30-second sliding window; the comfort evaluation channel records the cumulative value of the thermal discomfort index; the energy consumption monitoring channel statistics the total power consumption of the compressor; the safety compliance channel tracks the number of safety constraint violations. The fusion controller uses a feature concatenation connection method, and generates a value evaluation vector through a three-layer 256-node fully connected network. The strategy network gradient update executes a double-buffering mechanism: the online network collects experience samples every 15 milliseconds and stores them in the replay buffer; the target network performs soft update operations every 3000 steps, with a smoothing coefficient τ=0.005.

[0088] The control chain path topology verification module constructs an instruction directed graph. Each temperature regulation instruction is modeled as a graph node, with node attributes including: target temperature value (resolution 0.01°C), execution time window (accuracy 0.1 seconds), and clinical rule association identifier (128-bit mask). Dependency edges are divided into two types: sequential dependency edges use a single solid line to represent a forced sequence relationship; resource mutual exclusion edges use a double-dotted line to represent a device conflict relationship. The dependency graph is allowed to contain a maximum of 256 nodes, and when the threshold is exceeded, the graph partitioning algorithm is automatically started.

[0089] The back-reference mapping system uses a three-level index structure. The main index table stores the mapping relationship between the temperature instruction parameter combination and the clinical rule address, with a hash bucket depth of 16. The conflict detector performs seven rule verifications when traversing each instruction: whether the target temperature exceeds the 34.0-40.0°C boundary; whether the single adjustment temperature difference exceeds 0.8°C; whether the minute-level temperature change rate is higher than 0.5°C / min; whether the cumulative temperature difference of consecutive adjustment instructions exceeds 3°C; whether the maximum power of the device exceeds 2200W; whether the compressor continuous working time is lower than the safety threshold; whether the water flow rate is within the 10-300ml / min tolerance range. The violation identifier coding rule is: each binary flag corresponds to a specific rule violation state, forming a 16-bit conflict code.

[0090] Loop risk detection module implements optimized depth-first search. Graph traversal sets maximum depth of 100 layers, and stack space pre-allocates 16 KB memory. The search process maintains two auxiliary data structures: a real-time access flag array records node access status; a loop path cache stores node sequences that may form a loop. When the length of the detected instruction sequence exceeds 12, parallel search is started: the main thread executes standard depth-first traversal; the auxiliary thread verifies key node dependencies using a breadth-first strategy. When a potential loop is detected, the reconstruction engine is triggered: first, a virtual barrier node is added to cut off the loop; when a complex loop exists, a sequence reconstruction algorithm is started to regenerate a loop-free topology sequence. The final verified control instruction sequence is marked with a generation timestamp, a cyclic redundancy check code (32-bit polynomial), and a digital signature (based on an elliptic curve algorithm), forming a complete secure control instruction data packet.

[0091] Example 4: Focus on clinical safety rule verification and instruction sequence reconstruction mechanism. The clinical safety rule engine deploys a double-processing channel: the forward reasoning channel monitors physical parameters in real time, and the backward verification channel processes control instructions in batches. The body temperature safety boundary rule divides the 34-42°C interval into 32 subintervals with a width of 0.25°C, and each subinterval sets a hazard level score. The interval below 34°C activates the hypothermia risk flag, and the score increases linearly as the temperature decreases; the interval above 39°C triggers the high fever risk flag, and the score increases by 40%. The rule engine maintains a hazard level query table that dynamically updates the hazard threshold based on real-time physical data, with an update frequency of up to 5 times per second.

[0092] The temperature change rate threshold rule is implemented as a dual detection system. The main detector uses a 60-second wide ring buffer to store recent temperature values, and calculates the change rate acceleration parameter every second. When the standard deviation in the detection window exceeds the set value of 0.15°C, the auxiliary detection channel is automatically enabled: this channel is configured with a more sensitive 40-second moving window and a first-order derivative analysis to identify abnormal fluctuations. The device power limit rule is implemented through a current transformer and a voltage sampling circuit for real-time monitoring at the millisecond level, establishing a power-temperature correspondence mapping table. This mapping table is completed during the initial calibration of the device preheating stage, recording the typical power consumption curve of the compressor at different temperature set values.

[0093] Table 1: Safety rule conflict detection parameters.

[0094]

[0095] The conflict detector initiates a seven-step analysis procedure when performing instruction validation. Take the temperature regulation instruction with number #207 as an example: this instruction plans to decrease the ice blanket temperature from 35.8°C to 34.2°C within 180 seconds, containing three critical parameters. The rule engine first parses the target temperature 34.2°C, querying that the value lies in the 8th hazard interval (34.0-34.25°C), corresponding to the low temperature hazard level 0.78. Then it calculates the temperature difference change rate: (35.8-34.2) / 3=0.533°C / min, exceeding the 0.5°C / min threshold. Finally, it predicts the compressor power requirement: querying the mapping table finds that this working condition requires 2310W power, exceeding the 2200W rated value. The detector generates the conflict code 0x1A (binary 011010), marking the simultaneous triggering of the 1st, 3rd, and 5th rules in violation.

[0096] The instruction sequence reconfiguration engine initiates a topological sorting procedure. The initial instruction sequence contains 89 nodes, and a depth-first scan discovers a loop: node E (decrease temperature to 36.0°C) depends on node G (preheat to 38.0°C), while node G needs node E to complete before execution. The reconfiguration process first deletes the hot node #207 that violates the three rules; then it enables the topological sorting layering grouping algorithm: the remaining 86 nodes are divided into 12 levels, and nodes in the same level are allowed to execute in parallel. The level division is based on instruction time window overlap analysis: if two instructions have more than 30% overlap in execution time and the target temperature difference is greater than 1.0°C, they are classified into different levels.

[0097] The device initialization module adds two types of pre-instructions: the compressor preheat instruction contains three stages: the first 30 seconds increase to 32°C at a rate of 0.3°C / s; then maintain constant temperature for 120 seconds; finally adjust to the initial working temperature of 35.5°C. The sensor calibration instruction activates the closed-loop test procedure: apply three standard temperature points (30.0°C, 35.0°C, 40.0°C) for stimulation, and collect the response curve of each sensor for 30 seconds. The calibration data analyzer calculates the offset compensation value, for example, a forehead sensor detects 35.3°C when the reference temperature is 35.0°C, then write a -0.3°C compensation parameter.

[0098] The safety control instruction sequence generation adopts a frame structure. Each frame of data contains 256 bytes of payload: the first 8 bytes are the instruction sequence number and version identifier; then 96 bytes store the topologically layered instruction groups (12 bytes per instruction); followed by 64 bytes of device initialization parameters; the last 32 bytes contain safety check data. The frame structure is output through a time-triggered mechanism: a 5 millisecond communication window is reserved in the compressor operation cycle, which occurs every 200 milliseconds. The data sender injects 3 frames of data when the communication window is opened, forming a data pipeline.

[0099] The digital signature system adopts a hierarchical key architecture. The frame-level signature uses a 256-bit elliptic curve algorithm, and the private key is stored in a hardware security module. Take a real instruction frame as an example: the frame data with sequence number #8912 generates a digest value of 0x7A3F, and the signature generator encrypts the digest using the device-specific private key, outputting a 512-bit signature code. The system-level certificate binds the hospital ID and device sequence number, and the certificate renewal period is 24 hours, with an automatic renewal process starting 30 minutes before expiration. The transmission layer establishes a redundant check channel: when the main channel transmits instruction data through the CAN bus, the backup channel transmits instruction sequence numbers and temperature setting value parameters through the infrared link, realizing cross-channel data verification.

[0100] Example 5: The precise control and real-time feedback mechanism of the ice blanket machine temperature execution system is developed. The temperature instruction parser adopts a dual-mode working architecture, the standard mode directly reads the target parameters in the safety control instruction sequence, and the safety mode additionally activates the digital signature verification and redundant channel comparison program. When the instruction decoding unit processes each frame of 256 bytes of data, it first extracts the target temperature value field, which is stored as a 16-bit unsigned integer with a resolution of 0.01°C and an effective range of 30.00°C to 45.00°C. The execution time window parameters include a start timestamp and a duration of two 32-bit values, with a time reference synchronized to a medical-grade atomic clock and a minimum scheduling unit of 10 milliseconds. When the time window span exceeds 300 seconds, it is automatically divided into multiple sub-tasks for batch execution.

[0101] The compressor power regulation system is composed of three parallel controllers. The main controller implements an improved proportional-integral-derivative algorithm with a sampling period of 50 milliseconds, and outputs a pulse width modulation signal to drive the insulated gate bipolar transistor. The auxiliary controller runs a fuzzy logic algorithm, dynamically adjusts the control parameters based on recent temperature trends, and updates at a frequency of 20 times per second. The safety monitor continuously tracks the compressor coil temperature and the state of the heat sink, and immediately intervenes in control when overheating is detected. The power output stage uses four-quadrant control technology, allowing 0.1% precision forward refrigeration and reverse heating adjustment. In a typical working state, the compressor decreases from 35.0°C to 34.0°C in 90 seconds, during which the power output shows a parabolic feature of first increasing and then decreasing.

[0102] The circulating water flow control system is configured with a high-precision stepper motor driven valve group. The valve position controller receives a 0-100% opening instruction, and each 1% opening corresponds to 200 micro-steps of the stepper motor. The actual water flow calibration curve is stored in a non-volatile memory. The flow feedback system includes a Hall effect sensor and a turbine flowmeter double measurement channel, and a data fusion processor outputs a corrected flow value at a frequency of 250 Hz. When executing a cooling instruction, the valve is first quickly opened to an estimated position (such as from 30% to 65%), and then enters a fine tuning stage, with each adjustment not exceeding 2%. The warming process uses the opposite strategy, with conservative adjustment at the beginning and accelerated compensation at the end, to avoid sudden changes in water temperature causing patient discomfort.

[0103] The real-time feedback system establishes a multi-level data acquisition network. The main channel transmits device operating parameters through a controller area network bus, including 12 types of data such as actual compressor power, water temperature instantaneous value, flowmeter reading, etc., and sends a complete status frame every 100 milliseconds. The backup channel transmits a simplified data set using Bluetooth Low Energy protocol, including only key temperature values and abnormal flags, and the transmission interval can be dynamically adjusted, with a minimum of 50 milliseconds. The data synchronizer compares the timestamp difference of the two channel transmissions, and triggers the clock calibration program when the deviation exceeds 20 milliseconds.

[0104] The deviation analysis module performs a three-level verification process. The first level of verification calculates the instantaneous difference between the target temperature and the actual temperature, with an accuracy of 0.01°C. The second level of verification analyzes the cumulative temperature difference integral within a 300-second sliding window to identify systematic deviation trends. The third level of verification combines environmental temperature and patient body surface characteristics to predict the theoretical temperature difference allowed range. When the deviation exceeds the threshold for three consecutive samples, the system automatically responds in stages: 0.3°C deviation triggers parameter fine tuning; 0.5°C deviation starts control algorithm recalculation; 1.0°C or more deviation executes emergency protocol, suspends the current instruction and reinitializes the control system.

[0105] Data transmission optimization uses differential encoding technology. Temperature data stream uses 35.0°C as the reference value, and subsequent transmission only records the change amount, with a data compression rate of 60%. The timestamp uses relative encoding, with the first frame transmitting the complete time and subsequent frames transmitting millisecond-level offsets. The data packet structure is carefully designed: the header includes a 4-byte synchronization word and a 2-byte length identifier; the payload is packaged in TLV format; and the trailer adds a 32-bit cyclic redundancy check code. The network scheduler manages multiple device communication time slots to ensure that the transmission of the ice blanket machine, the monitor and the central control console does not interfere with each other, with a maximum end-to-end delay of 80 milliseconds.

[0106] The exception handling mechanism contains five levels of recovery. Level one exception is a transient communication disruption, the system automatically switches to the backup channel and retransmits the last three frames of data. Level two exception involves a single sensor failure, the adjacent sensor data replacement algorithm is started. Level three exception manifests as actuator response delay, triggering a timeout retry mechanism with a maximum of three attempts. Level four exception corresponds to a severe parameter deviation, the safety protocol aborts the current operation and falls back to the last stable state. Level five exception is a multiple system failure, the actuator power is immediately cut off, the audible and visual alarms are activated and an emergency interrupt signal is sent to the medical monitoring system. Each exception level corresponds to a specific recovery script, stored in a tamper-proof memory, which can be forcibly invoked through a physical switch.

[0107] The device status visualization interface presents a multi-layer information structure. The base layer displays the current temperature setpoint, actual value and change curve, with a refresh rate of 10 frames per second. The middle layer shows device parameters such as compressor power percentage, water flow rate and remaining operating time. The advanced layer requires authorization to unlock, providing diagnostic information such as control algorithm internal state, network quality indicators and exception logs. All display elements are optimized through human factors engineering, with key parameters using high-contrast color coding, automatically magnified and displayed with dynamic borders when the value changes beyond a threshold. The operation log is stored in a ring buffer structure, recording the last 1000 important events, including instruction execution status, mode switching time and exception handling records, supporting quick retrieval by time range or event type.

[0108] It should be noted that the relative terms, such as first and second, and the like, are used herein solely to distinguish one from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0109] While embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, combinations, and variations can be made by those skilled in the art without departing from the spirit and scope of the present application, which is defined by the appended claims and their equivalents.

Claims

1. A dynamic temperature control method for an ice blanket machine based on body temperature feedback, characterized in that, The method comprises the following steps: Body temperature multi-source perception, obtaining target object's body surface temperature perception data and core temperature simulation data, and constructing original body temperature perception data set; Physiological state evaluation, generating physiological state feature vector according to the body surface temperature perception data and the core temperature simulation data; Multi-modal body temperature state modeling, generating multi-modal body temperature state tensor through multi-source perception fusion modeling based on the physiological state feature vector; Body temperature regulation strategy optimization, generating optimized temperature control strategy by using reinforcement learning method with dynamic binding of clinical constraint rule component according to the multi-modal body temperature state tensor; Control chain path topology verification, performing integrity verification and clinical safety rule conflict detection on the optimized temperature control strategy to generate safe control instruction sequence; Dynamic temperature execution, driving ice blanket machine to perform temperature regulation operation according to the safe control instruction sequence; The body temperature multi-source perception comprises the following operations: Collecting skin surface temperature time sequence through distributed temperature sensor, and synchronously acquiring environmental temperature and humidity compensation parameters; Generating core temperature simulation value through core temperature estimation algorithm, combining the skin surface temperature time sequence and the environmental temperature and humidity compensation parameters, and constructing original body temperature perception data set; The physiological state evaluation comprises the following operations: Performing temperature time sequence correction on the original body temperature perception data set to generate body surface temperature time sequence; Inputting the body surface temperature time sequence into physiological state mapping component to generate core temperature offset feature through body surface-core temperature correlation model; Fusing the core temperature offset feature and the environmental temperature and humidity compensation parameters to generate physiological state feature vector; The multi-modal body temperature state modeling comprises the following operations: Constructing hierarchical attention fusion mechanism, including body temperature regulation response attention layer, environmental interference suppression attention layer and clinical risk early warning attention layer; Extracting temperature regulation demand feature in the physiological state feature vector through the body temperature regulation response attention layer; Filtering environmental temperature and humidity interference noise feature through the environmental interference suppression attention layer; Generating body temperature sudden change risk probability distribution through the clinical risk early warning attention layer; Aggregating the temperature regulation demand feature, the environmental temperature and humidity interference noise feature and the body temperature sudden change risk probability distribution to generate multi-modal body temperature state tensor.

2. The body temperature feedback based ice blanket machine dynamic temperature control method of claim 1, wherein, The body temperature regulation strategy optimization comprises the following operations: Constructing reinforcement learning state parameter space, including ice blanket current temperature state, body temperature change rate state and clinical constraint rule component state; Using multi-layer strategy network architecture, including annual scheme generation layer, clinical dynamic optimization layer and real-time execution decision layer; Outputting long-term body temperature management framework protocol through the annual scheme generation layer; Generating hourly temperature control scheme by dynamically adjusting the long-term body temperature management framework protocol through the clinical dynamic optimization layer; Generating minute-level temperature regulation instruction through the real-time execution decision layer.

3. The body-temperature feedback based ice blanket machine dynamic temperature control method of claim 2, wherein, The training of the multi-layer strategy network architecture comprises the following operations: Designing multi-objective reward aggregation function, integrating temperature offset penalty factor, body discomfort evaluation factor and device energy consumption factor; Introducing clinical disturbance simulation function to simulate body temperature abnormal fluctuation event, sensor failure event and environmental mutation event; The optimization temperature control strategy is generated by a staged optimization process, including an initial protocol generation stage, a clinical disturbance adaptation stage, and a multi-objective strategy aggregation stage.

4. The body-temperature feedback based ice blanket machine dynamic temperature control method of claim 3, wherein, The control chain path topology verification includes the following operations: A control instruction dependency graph is constructed to record the execution dependency relationship of each temperature regulation instruction in the optimization temperature control strategy; Conflict nodes of the temperature regulation instructions and the clinical safety rule components are detected by a reverse reference mapping table; A deep-first search algorithm is used to verify the loop risk of the control instruction dependency graph; A safe control instruction sequence without conflict and loop is generated.

5. The body-temperature feedback based ice blanket machine dynamic temperature control method of claim 4, wherein, The conflict detection of the clinical safety rule components includes the following operations: A body temperature safety boundary rule, a temperature change rate threshold rule, and a device maximum power limit rule are defined; Each temperature regulation instruction in the optimization temperature control strategy is traversed, and violation identifiers of the body temperature safety boundary rule, the temperature change rate threshold rule, and the device maximum power limit rule are matched by a bidirectional mapping tool; Temperature regulation instructions containing the violation identifiers are deleted.

6. The ice blanket machine dynamic temperature control method based on body temperature feedback of claim 5, wherein, The generation of the safe control instruction sequence includes the following operations: The conflict-free temperature regulation instructions are sorted according to the execution dependency relationship based on a topological sorting algorithm; A device startup preheating instruction and a sensor calibration instruction are added as pre-node; A safe control instruction sequence in sequence is output.

7. The body-temperature feedback based ice blanket machine dynamic temperature control method of claim 6, wherein, The dynamic temperature execution includes the following operations: Temperature target values and execution time windows in the safe control instruction sequence are parsed; An ice blanket compressor power regulation component and a circulating water flow control valve are driven to gradually approach the temperature target values according to the execution time windows; Real-time feedback of the deviation of the actual temperature value from the temperature target value to the body temperature multi-source perception step.

Images