Method for compensating internal thermal drift of marine multi-parameter precision measuring instrument

Abstract

The application provides a method for internal thermal drift compensation of a marine multi-parameter precision measuring instrument, and belongs to the technical field of marine equipment.The application dynamically adjusts the weight parameters of a thermal field topology compensation model according to a comprehensive adjustment factor interval through a self-adaptive adjustment function, and then outputs a feedforward compensation value through a bidirectional long short-term memory network and a self-adaptive Kalman filter and fuses the feedforward compensation value with a thermal drift prediction value to realize online thermal drift correction.Meanwhile, an incremental learning mechanism is used to perform online adaptive weight adjustment on the dynamic offset of a thermal drift reference line.Finally, laboratory calibration and consistency evaluation are used to perform feedback correction on local thermal capacity parameters.Through the collaborative mechanism of artificial intelligence and physical modeling, the technical problem that the marine multi-parameter precision measuring instrument cannot realize online adaptive compensation due to the systematic drift of measurement data caused by the dynamic change of the internal thermal field under the condition of multi-channel synchronous sampling is solved.

Description

Technical Field

[0001] This invention belongs to the field of marine equipment technology, and more specifically, relates to a method for compensating for internal thermal drift in a marine multi-parameter precision measuring instrument. Background Technology

[0002] Marine multi-parameter precision measuring instruments are widely used in hydrological surveys, marine environmental monitoring, and deep-sea engineering exploration. These instruments typically integrate multiple high-precision sensing channels for temperature, salinity, pressure, and dissolved oxygen, with each channel sampling synchronously and outputting measurement data in real time. In current engineering applications of marine multi-parameter precision measuring instruments, the presence of multiple heterogeneous heat sources within the instrument, coupled with uneven power distribution and internal / external temperature gradients caused by drastic changes in external water temperature, results in a highly dynamic and unsteady-state thermal field within the instrument, both in time and space. This leads to continuous drift of the measurement parameters of each sensing channel as the internal temperature changes. Existing methods primarily rely on low-temperature drift device selection, static calibration correction, or offline thermal compensation models. These methods have significant limitations in handling multi-channel coupled thermal drift. Static models cannot track the dynamic shift of the thermal drift baseline caused by device aging, and offline compensation cannot respond to transient thermal disturbances during actual instrument operation. In other words, existing technologies present a technical problem: under multi-channel synchronous sampling conditions, marine multi-parameter precision measuring instruments experience systematic data drift due to dynamic changes in the internal thermal field, and online adaptive compensation is not feasible. Summary of the Invention

[0003] In view of this, the present invention provides a method for internal thermal drift compensation of a marine multi-parameter precision measuring instrument, which can solve the technical problem in the prior art where the measurement data of a marine multi-parameter precision measuring instrument undergoes systematic drift due to dynamic changes in the internal thermal field under multi-channel synchronous sampling conditions and cannot achieve online adaptive compensation.

[0004] This invention is implemented as follows: This invention provides a method for compensating for internal thermal drift in a marine multi-parameter precision measuring instrument, comprising the following steps:

[0005] Multiple thermistors are deployed inside the marine multi-parameter precision measuring instrument. A sliding window streaming processing architecture combined with an asynchronous coroutine mechanism is used to schedule the multi-channel high-frequency temperature sampling data streams output by each thermistor. Hash feature dimensionality reduction is used to remove redundant features from the multi-channel high-frequency temperature sampling data streams to construct an internal temperature distribution matrix.

[0006] The partial differential equation of thermal fluid convection in the inner cavity of the marine multi-parameter precision measuring instrument is extracted and transformed into a multidimensional temperature field state transition matrix. The real-time temperature readings of each thermosensitive element in the internal temperature distribution matrix are used as input variables. The extreme point of the energy dissipation objective function is found in the thermal resistance network topology using the fluid thermodynamic gradient descent fusion algorithm. The thermal drift compensation coefficient of each measurement channel is output.

[0007] The internal temperature distribution matrix and the measurement parameters collected by each measurement channel are input into the thermal field topology compensation model. The thermal field topology compensation model outputs the thermal drift prediction value of each measurement channel. The deviation between the thermal drift prediction value and the current measurement parameter of each measurement channel is calculated. The deviation is input into the adaptive adjustment function. The weight parameters of the thermal field topology compensation model are dynamically adjusted according to the interval of the comprehensive adjustment factor output by the adaptive adjustment function.

[0008] A bidirectional long short-term memory network is used to process the temperature time series data in the internal temperature distribution matrix, capture the dynamic hysteresis characteristics of the temperature gradient, and output the feedforward compensation value. An adaptive Kalman filter is introduced to filter the transient thermal noise of the feedforward compensation value. The filtered feedforward compensation value is fused with the thermal drift prediction value to perform online thermal drift correction on the measurement parameters of each measurement channel and output the corrected measurement data.

[0009] During the continuous operation of the marine multi-parameter precision measuring instrument, the dynamic offset of the thermal drift baseline of each measuring channel is monitored. The dynamic offset is input into the incremental learning mechanism, and the thermal field topology compensation model is adjusted online only for the dynamic offset to maintain the long-term effectiveness of the thermal field topology compensation model.

[0010] During the laboratory calibration phase, multiple temperature gradient excitations were applied to the marine multi-parameter precision measuring instrument, and the thermal drift amplitude of each measurement channel within the typical operating temperature range was recorded. The consistency between the calibration data and the corrected measurement data was evaluated, and the local thermal capacity parameters in the thermal resistance network topology were adjusted based on the consistency evaluation results. The adjusted local thermal capacity parameters were then fed back to the fluid thermodynamic gradient descent fusion algorithm.

[0011] Specifically, the sliding window streaming architecture refers to segmenting and batch processing multi-channel high-frequency temperature sampling data streams using a fixed-length time window, and using an asynchronous coroutine mechanism to achieve concurrent scheduling of data streams from each channel, thereby avoiding thread blocking.

[0012] Specifically, the hash feature dimensionality reduction refers to performing hash mapping on multidimensional temperature feature vectors, compressing high-dimensional features into a low-dimensional space, retaining the main thermal drift information, and eliminating redundant temperature features.

[0013] The multidimensional temperature field state transition matrix describes the transmission relationship of the cavity temperature field between time steps. The temperature change of a thermistor over time is determined by the adjacency coefficient of the thermal resistance network topology, the power of the local heat source, and the local heat capacity. Each parameter is normalized to its corresponding reference value before being used in the calculation.

[0014] Specifically, the fluid thermodynamic gradient descent fusion algorithm refers to using the real-time temperature readings of each thermistor as the state variable and energy dissipation as the objective function in the thermal resistance network topology. In each iteration, the learning rate step size is dynamically adjusted according to the local heat capacity parameters, and after multiple rounds of convergence, the thermal drift compensation coefficients of each measurement channel are output.

[0015] The thermal field topology compensation model adopts a graph convolutional neural network structure, treating the thermal element as a graph node and the heat conduction path as the edge of the graph. It performs information aggregation and message passing iteration between graph nodes through the adjacency matrix, uses skip connections to prevent feature over-smoothing, and outputs the thermal drift prediction value of each measurement channel.

[0016] The establishment of the training dataset for the thermal field topology compensation model specifically includes: applying different temperature gradient excitations to the marine multi-parameter precision measuring instrument, synchronously recording the real-time temperature readings of each thermosensitive element, the measurement parameters of each measurement channel and the corresponding calibration thermal drift true values, and constructing temperature-drift sample pairs under multiple operating conditions.

[0017] Specifically, the training of the thermal field topology compensation model uses the mean square error between the predicted thermal drift value and the calibrated true thermal drift value as the main term of the loss function. The constraints of the heat conduction law are embedded into the loss function in the form of a physical information neural network. The weight parameters of the thermal field topology compensation model are iteratively optimized using the mini-batch gradient descent method until the loss function converges.

[0018] The adaptive adjustment function calculates the comprehensive adjustment factor using the mean deviation, the rate of change of deviation, and the confidence index as inputs. When the comprehensive adjustment factor is not less than 1.5, a large step weight update strategy is adopted; when the comprehensive adjustment factor is in the interval [0.8, 1.5), a standard step weight update strategy is adopted; when the comprehensive adjustment factor is in the interval [0.3, 0.8), a small step weight update strategy is adopted; and when the comprehensive adjustment factor is less than 0.3, weight update is paused.

[0019] The confidence index refers specifically to a quantitative indicator of the stability of the thermal drift prediction output of the thermal topology compensation model under the current input conditions, which is calculated from the normalized activation variance of the output layer of the thermal topology compensation model.

[0020] The bidirectional long short-term memory network models the temperature time series data in the internal temperature distribution matrix in both forward and reverse directions, while capturing the dynamic hysteresis characteristics during the temperature gradient rise and fall process and outputting feedforward compensation values.

[0021] Specifically, the adaptive Kalman filter is based on the standard Kalman filter, and dynamically adjusts the filter gain according to the real-time estimation of the measurement noise covariance to filter the transient thermal noise introduced by the drastic change in water temperature in the feedforward compensation value.

[0022] Specifically, the thermal drift baseline refers to the slow, monotonic drift trend of each measurement channel under conditions without external excitation or disturbance, reflecting the thermal drift zero-point offset caused by device aging.

[0023] Specifically, the incremental learning mechanism refers to updating the local weight parameters of the thermal field topology compensation model using only the newly added samples corresponding to the dynamic offset, without reconstructing all weight parameters, thereby reducing the computational cost of retraining while suppressing the forgetting of historical knowledge.

[0024] Specifically, the consistency assessment involves statistically analyzing the mean square error of the calibration data and the corrected measurement data under multiple temperature gradient excitation conditions, and using whether the mean square error is lower than a preset threshold as the basis for determining whether the local heat capacity parameters in the thermal resistance network topology need to be adjusted.

[0025] This invention collects multi-dimensional temperature distribution data by deploying an array of thermistors inside the instrument. It then establishes a thermal resistance network topology model using a fluid thermodynamic gradient descent fusion algorithm. A thermal field topology compensation model is used to predict multi-channel thermal drift, and a bidirectional long short-term memory network and adaptive Kalman filtering are introduced to achieve feedforward compensation for dynamic thermal drift. Simultaneously, an incremental learning mechanism is used to adaptively update the thermal field topology compensation model online. This solves the technical problem of systematic data drift caused by dynamic changes in the internal thermal field under multi-channel synchronous sampling conditions, which prevents online adaptive compensation. This invention explicitly models the physical conduction laws of the internal thermal field using a graph convolution structure, enabling the compensation model to perceive the spatial propagation characteristics of the thermal field and accurately predict the thermal drift of each channel when thermal disturbances occur. By embedding the constraints of the thermal conduction law into the loss function, the model is constrained by prior physical knowledge under finite sample conditions, resulting in faster convergence and enhanced generalization ability. In summary, this invention solves the technical problem mentioned in the background art of systematic data drift caused by dynamic changes in the internal thermal field under multi-channel synchronous sampling conditions in marine multi-parameter precision measuring instruments, which prevents online adaptive compensation. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 The time-series temperature distribution curves inside each thermistor are shown.

[0028] Figure 3This is to trigger an interval graph by comprehensively adjusting the dynamic changes of the adjustment factor and the weight update strategy.

[0029] Figure 4 A comparison curve of thermal drift amplitude before and after compensation for each measurement channel. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0031] like Figure 1 The diagram shown is a flowchart of a method for compensating for internal thermal drift in a marine multi-parameter precision measuring instrument provided by the present invention. This method includes the following steps:

[0032] S01. Multiple thermal elements are deployed inside the marine multi-parameter precision measuring instrument. The multi-channel high-frequency temperature sampling data stream output by each thermal element is scheduled using a sliding window streaming processing architecture combined with an asynchronous coroutine mechanism. Redundant features of the multi-channel high-frequency temperature sampling data stream are removed by hash feature dimensionality reduction to construct an internal temperature distribution matrix.

[0033] S02. Extract the partial differential equation of thermal fluid convection in the inner cavity of the marine multi-parameter precision measuring instrument, transform the partial differential equation of thermal fluid convection into a multidimensional temperature field state transition matrix, take the real-time temperature readings of each thermosensitive element in the internal temperature distribution matrix as input variables, find the extreme point of the energy dissipation objective function in the thermal resistance network topology using the fluid thermodynamic gradient descent fusion algorithm, and output the thermal drift compensation coefficient of each measurement channel.

[0034] S03. Input the internal temperature distribution matrix and the measurement parameters collected by each measurement channel into the thermal field topology compensation model. The thermal field topology compensation model outputs the thermal drift prediction value of each measurement channel. Calculate the deviation between the thermal drift prediction value and the current measurement parameter of each measurement channel. Input the deviation into the adaptive adjustment function. Dynamically adjust the weight parameters of the thermal field topology compensation model according to the interval of the comprehensive adjustment factor output by the adaptive adjustment function.

[0035] S04. A bidirectional long short-term memory network is used to process the temperature time series data in the internal temperature distribution matrix, capture the dynamic hysteresis characteristics of the temperature gradient, and output the feedforward compensation value. An adaptive Kalman filter is introduced to filter the transient thermal noise of the feedforward compensation value. The filtered feedforward compensation value is fused with the thermal drift prediction value to perform online thermal drift correction on the measurement parameters of each measurement channel and output the corrected measurement data.

[0036] S05. During the continuous operation of the marine multi-parameter precision measuring instrument, monitor the dynamic offset of the thermal drift baseline of each measuring channel, input the dynamic offset into the incremental learning mechanism, and perform online adaptive weight adjustment on the thermal field topology compensation model only for the dynamic offset to maintain the long-term effectiveness of the thermal field topology compensation model.

[0037] S06. In the laboratory calibration stage, multiple sets of temperature gradient excitations are applied to the marine multi-parameter precision measuring instrument, and the thermal drift amplitude of each measurement channel within the typical operating temperature range is recorded. The consistency evaluation of the calibration data and the corrected measurement data is carried out. Based on the consistency evaluation results, the local thermal capacity parameters in the thermal resistance network topology are adjusted, and the adjusted local thermal capacity parameters are fed back to the fluid thermodynamic gradient descent fusion algorithm in S02.

[0038] The sliding window streaming architecture refers to a data scheduling method that segments and buffers multi-channel high-frequency temperature sampling data streams and processes them in batches using a fixed-length time window. It is combined with an asynchronous coroutine mechanism to achieve concurrent scheduling of data streams from each channel and avoid thread blocking.

[0039] The hash feature dimensionality reduction refers to performing hash mapping on multidimensional temperature feature vectors, compressing high-dimensional features into a low-dimensional space, retaining the main thermal drift information, eliminating redundant temperature features, and reducing the interference of redundant features on subsequent modeling.

[0040] The measurement parameters include measured values ​​from four types of sensor channels: temperature, salinity, pressure, and dissolved oxygen.

[0041] The multidimensional temperature field state transition matrix describes the transmission relationship of the internal cavity temperature field between time steps, and the formula is expressed as follows: ;in For the first Temperature of each thermistor element, in °C. Reference temperature, unit: °C The time step is in seconds. For reference time, the unit is seconds. The adjacency coefficient of the thermal resistance network topology is dimensionless. This refers to the power of a local heat source, measured in W. For reference heat source power, the unit is W. This is the local heat capacity, in units of For reference heat capacity, the unit is... .

[0042] The fluid thermodynamic gradient descent fusion algorithm refers to using the real-time temperature readings of each thermistor as the state variable and energy dissipation as the objective function in a thermal resistance network topology. In each iteration, the learning rate step size is dynamically adjusted based on local heat capacity parameters. After multiple rounds of convergence, the thermal drift compensation coefficients for each measurement channel are output. The learning rate adjustment formula is as follows: ;in For the first The learning rate for each iteration, in units of 1. This is the initial learning rate, in units of 1. For reference learning rate, the unit is 1. The attenuation factor is dimensionless. For the first The local heat capacity parameter at the next iteration, in units of For reference heat capacity, the unit is... .

[0043] The thermal field topology compensation model employs a graph convolutional neural network structure, treating the thermosensitive element as a graph node and the heat conduction path as the graph edge. It iterates information aggregation and message passing between graph nodes using an adjacency matrix, utilizes skip connections to prevent feature oversmoothing, extracts global thermal distribution features, and outputs the predicted thermal drift values ​​for each measurement channel. The specific structure of the thermal field topology compensation model is as follows: the input layer receives the internal temperature distribution matrix and the measurement parameters of each measurement channel; multiple graph convolutional layers sequentially perform neighborhood aggregation on the graph node features; each graph convolutional layer is followed by a residual skip connection; and the output layer outputs the predicted thermal drift values ​​for each measurement channel. The steps for establishing the training dataset for the thermal field topology compensation model specifically include: applying different temperature gradient excitations to a marine multi-parameter precision measuring instrument in a laboratory environment, simultaneously recording the real-time temperature readings of each thermosensitive element, the measurement parameters of each measurement channel, and the corresponding calibrated true thermal drift values, constructing temperature-drift sample pairs under multiple operating conditions to form the training dataset. The specific steps for training the thermal field topology compensation model include: using the mean square error between the predicted thermal drift value and the calibrated true thermal drift value as the main term of the loss function; embedding the thermal conduction law constraint into the loss function in the form of a physical information neural network; and iteratively optimizing the weight parameters of the thermal field topology compensation model using the mini-batch gradient descent method until the loss function converges.

[0044] The described thermal field topology compensation model explicitly models the heat conduction relationship between nodes of the thermal element using a graph convolution structure. This allows the model to incorporate the physical propagation laws of the internal thermal field into the feature extraction process when extracting thermal drift features. By embedding the thermal conduction law constraint into the loss function, the model is guided by prior physical knowledge during training, enabling it to converge quickly even with limited training samples and suppressing overfitting. Skip connections preserve shallow local thermal distribution features, which, together with deep global features, allow the model to simultaneously perceive local heat accumulation and overall thermal field changes. This enables accurate prediction of thermal drift across multiple measurement channels, improving measurement stability and long-term reliability without increasing hardware costs.

[0045] The adaptive adjustment function takes the mean deviation between the current thermal drift prediction value and the measured parameters, the rate of change of the deviation, and the confidence index output by the thermal field topology compensation model as inputs to calculate the comprehensive adjustment factor, as expressed in the following formula: ;in As a comprehensive regulatory factor, dimensionless This is the mean deviation, in units of 1. This is a reference deviation, in units of 1. The deviation rate is expressed in units of... The rate of change of reference deviation, in units of The confidence index is dimensionless. For reference confidence level, dimensionless. , , is the weighting coefficient, dimensionless. When When a large-step weight update strategy is used, the weight parameters of the thermal field topology compensation model are quickly adjusted; when When a standard step-size weight update strategy is used, the weight parameters of the thermal field topology compensation model are adjusted routinely; when At that time, a small-step weight update strategy is used to finely adjust the weight parameters of the thermal field topology compensation model; when At this time, pause the weight update and keep the weight parameters of the thermal field topology compensation model unchanged.

[0046] The confidence index refers to a quantitative indicator of the stability of the thermal drift prediction output of the thermal topology compensation model under the current input conditions, and is calculated from the normalized activation variance of the output layer of the thermal topology compensation model.

[0047] The bidirectional long short-term memory network models the temperature time series data in the internal temperature distribution matrix in both forward and reverse directions, while capturing the dynamic hysteresis characteristics during the temperature gradient rise and fall process and outputting feedforward compensation values.

[0048] The adaptive Kalman filter, based on the standard Kalman filter, dynamically adjusts the filter gain according to the real-time estimation of the measurement noise covariance, and filters the transient thermal noise introduced by the drastic change in water temperature in the feedforward compensation value, so as to ensure the stability of the filtered feedforward compensation value.

[0049] The thermal drift baseline refers to the slow, monotonic drift trend of each measurement channel under no external excitation disturbance during the long-term operation of the marine multi-parameter precision measuring instrument, reflecting the thermal drift zero-point offset caused by device aging.

[0050] The incremental learning mechanism refers to updating the local weight parameters of the thermal topology compensation model only using the newly added samples corresponding to the dynamic offset during the continuous operation of the thermal topology compensation model, without reconstructing all weight parameters, thereby reducing the computational cost of retraining while suppressing the forgetting of historical knowledge.

[0051] The consistency assessment refers to the statistical analysis of the mean square error between the calibration data and the corrected measurement data under multiple temperature gradient excitation conditions, and whether the mean square error is lower than a preset threshold is used as the basis for judging whether the local heat capacity parameters in the thermal resistance network topology need to be adjusted.

[0052] Optionally, the present invention also provides a method for implementing a marine multi-parameter precision measuring instrument internal thermal drift compensation system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions are used to execute the above-described method when the computer is run.

[0053] The specific implementation of step S01 is as follows: Multiple negative temperature coefficient thermistors are deployed at thermally sensitive locations such as the printed circuit board, sensor mounting base, and sealed cavity wall inside the marine multi-parameter precision measuring instrument. Each thermistor is connected to the data acquisition unit via an independent sampling channel, and the sampling frequency is set to no less than 10Hz to ensure the time resolution for dynamic changes in the internal thermal field. Since the temperature data stream generated by multi-channel high-frequency sampling is large, directly storing it in memory would exhaust microcontroller resources. Therefore, a sliding window streaming processing architecture is used to segment and cache the data stream. The window length reference value is 128 sampling points, and the window sliding step reference value is 32 sampling points. An asynchronous coroutine mechanism is used to achieve concurrent scheduling of the multi-channel data stream, avoiding data loss due to thread blocking. In the feature extraction stage, the temperature time-series data of each channel is converted into multi-dimensional feature vectors. A hash feature dimensionality reduction method is used to map the high-dimensional feature vectors to a low-dimensional space. The dimensionality reference value after dimensionality reduction is 30% of the original dimension. While retaining the main thermal drift information, redundant temperature features are removed. Finally, the dimensionality-reduced feature vectors are organized into an internal temperature distribution matrix for use in subsequent steps.

[0054] The specific implementation of step S02 is as follows: Based on the geometric structure and heat source distribution of the inner cavity of the marine multi-parameter precision measuring instrument, a partial differential equation for thermal fluid convection is established to describe the convective heat transfer process of the gas or liquid medium within the cavity. The partial differential equation for thermal fluid convection is discretized spatially using the finite difference method, transforming it into a multidimensional temperature field state transition matrix with thermistor nodes as state variables. The matrix dimension is consistent with the number of thermistors. Using the real-time temperature readings of each thermistor in the internal temperature distribution matrix as input variables, an energy dissipation objective function is constructed in the thermal resistance network topology. The extreme points are iteratively solved using a fluid thermodynamic gradient descent fusion algorithm. During each iteration, the learning rate step size of the gradient descent is dynamically adjusted based on the local heat capacity parameters in the thermal resistance network topology. The initial learning rate is referenced to be 0.01, and the decay factor is referenced to be 0.5. After multiple iterations, the change in the objective function is reduced to less than... When convergence is confirmed, the thermal drift compensation coefficients of each measurement channel are output for subsequent initialization and calibration of the thermal field topology compensation model.

[0055] The specific implementation of step S03 is as follows: The thermal field topology compensation model uses a graph convolutional neural network as its core structure. It takes the internal temperature distribution matrix and the measurement parameters collected from each measurement channel as input. The measurement parameters include the measured values ​​of four types of sensing channels: temperature, salinity, pressure, and dissolved oxygen. The model input layer concatenates the features of the thermal element nodes with the measurement parameter features and then feeds them into multiple graph convolutional layers. Each layer uses an adjacency matrix to weight and aggregate the features on the heat conduction path between nodes. Each graph convolutional layer is followed by a residual skip connection to prevent over-smoothing of features. The output layer outputs the predicted thermal drift values ​​for each measurement channel. After obtaining the predicted thermal drift values, the deviation between the predicted thermal drift values ​​and the current measurement parameters of each measurement channel is calculated. The mean deviation, the rate of change of deviation, and the confidence index are input into an adaptive adjustment function to calculate a comprehensive adjustment factor. When the comprehensive adjustment factor is not less than 1.5, a large step size weight update strategy is adopted to quickly adjust the thermal field topology compensation model, with the step size reference value being 3 times the baseline step size; when the comprehensive adjustment factor is in the interval [0.8, 1.5), a standard step size weight update strategy is adopted, with the step size being the baseline step size; when the comprehensive adjustment factor is in the interval [0.3, 0.8), a small step size weight update strategy is adopted, with the step size reference value being 0.3 times the baseline step size; when the comprehensive adjustment factor is less than 0.3, weight update is paused, and the current weight parameters of the thermal field topology compensation model remain unchanged.

[0056] The specific implementation of step S04 is as follows: Temperature time-series data from the internal temperature distribution matrix is ​​input into a bidirectional long short-term memory network. The network performs sequence modeling of the temperature time-series data in both forward and reverse directions. The forward channel captures the hysteresis characteristics of the temperature gradient increase process, while the reverse channel captures the recovery characteristics of the temperature gradient decrease process. The hidden states from both directions are concatenated and output as feedforward compensation values ​​through a fully connected layer. Since rapid water temperature changes introduce transient thermal noise into the feedforward compensation values, an adaptive Kalman filter is introduced to filter them. During the filtering process, the measurement noise covariance is estimated in real time based on the current temperature change rate, and the filter gain is dynamically adjusted. The filtered feedforward compensation values ​​are weighted and fused with the thermal drift prediction values ​​output by the thermal field topology compensation model. The fusion weights are determined based on the real-time confidence indices of the two signals. Finally, online thermal drift correction is performed on the measurement parameters of each measurement channel, and the corrected measurement data is output.

[0057] The specific implementation of step S05 is as follows: During the continuous operation of the marine multi-parameter precision measuring instrument, the corrected measurement data is continuously monitored using a sliding window streaming processing architecture. A linear trend estimation method is used to extract the dynamic offset of the thermal drift baseline for each measurement channel. When the change in dynamic offset exceeds a preset threshold within 24 consecutive hours, the newly added sample corresponding to the dynamic offset is input into the incremental learning mechanism. The incremental learning mechanism only locally updates the output layer weight parameters of the thermal field topology compensation model, with an update step size reference value of 0.1 times the baseline step size. It does not reconstruct all weight parameters of the convolutional layer, thereby tracking the dynamic offset of the thermal drift baseline while retaining historical compensation knowledge, achieving long-term adaptive effectiveness of the thermal field topology compensation model.

[0058] The specific implementation of step S06 is as follows: During the laboratory calibration stage, the marine multi-parameter precision measuring instrument is placed in a temperature-controlled environmental simulation chamber, and several sets of temperature gradient excitations are applied respectively. The reference value for the temperature change rate is 0.5~2℃ / min, and the temperature range covers the typical operating range of the instrument. Under each temperature gradient excitation condition, the thermal drift amplitude of each measurement channel is recorded synchronously as calibration data. The mean square error of the calibration data and the corrected measurement data under each set of temperature gradient excitation conditions is statistically analyzed, and consistency evaluation is performed with the mean square error being lower than a preset threshold of 0.05. If the consistency evaluation result shows that the mean square error exceeds the preset threshold, the local heat capacity parameters of each node in the thermal resistance network topology are adjusted according to the deviation distribution under each set of temperature gradient excitations, and the adjusted local heat capacity parameters are fed back to the fluid thermodynamic gradient descent fusion algorithm in step S02 to resolve the thermal drift compensation coefficient, forming a closed-loop calibration.

[0059] The key technical ideas and their effects of this invention are as follows. First, the fluid thermodynamic gradient descent fusion algorithm transforms the heat transfer process of the hot fluid inside the instrument cavity into an energy dissipation optimization problem in the thermal resistance network topology. By iteratively solving the extreme points within a physical constraint framework, the calculation process of the thermal drift compensation coefficient has a clear thermodynamic basis. Compared with purely data-based regression methods, it has stronger generalization ability and can maintain the physical consistency of the compensation coefficient even when the instrument's operating conditions change. Second, the thermal field topology compensation model uses a graph convolutional neural network to structurally model the thermal conduction topology relationship between the nodes of the thermal element. The thermal conduction law constraint is embedded in the loss function, so that the model is always guided by physical laws during the feature learning process. This avoids overfitting of the purely data-driven model in the sparse region of the thermal field features. At the same time, skip connections allow local thermal accumulation features and global thermal field features to participate in compensation prediction together. Third, the combination mechanism of bidirectional long short-term memory network and adaptive Kalman filter performs bidirectional modeling of the dynamic hysteresis process of thermal drift in the time dimension, and eliminates the interference of transient thermal noise through adaptive filtering, so that the feedforward compensation value remains stable and effective even in scenarios with drastic water temperature changes. The synergistic effect of the above three technical approaches is as follows: the fluid thermodynamic gradient descent fusion algorithm provides physically consistent initial compensation coefficients for the thermal field topology compensation model; the thermal field topology compensation model refines the spatial characteristics of multi-channel coupled thermal drift through graph structure learning on this basis; and the bidirectional long short-term memory network supplements the thermal field topology compensation model's lack of perception of dynamic hysteresis processes in the time dimension. The three complement each other at the three levels of physical modeling, spatial feature learning, and time series modeling, respectively, forming a complete multi-dimensional thermal drift compensation system.

[0060] It should be noted that this invention also solves the following technical problems: In the long-term operation of marine multi-parameter precision measuring instruments, device aging causes dynamic shifts in the thermal drift baseline, shortening the effective lifespan of traditional static compensation models and requiring frequent full recalibration, consuming significant computational and manual resources. This invention, through an incremental learning mechanism, updates the weight parameters locally only for newly added samples corresponding to the dynamic shift of the thermal drift baseline, without reconstructing all weight parameters of the thermal field topology compensation model. This continuously tracks the baseline shift while suppressing the forgetting of historical knowledge, solving the technical problem of the short lifespan of static compensation models. Furthermore, this invention also solves the technical problems of memory overflow and thread blocking when processing multi-channel high-frequency temperature sampling data streams under microcontroller resource constraints. It uses a sliding window streaming architecture to segment, cache, and batch process the data stream, combined with an asynchronous coroutine mechanism to achieve concurrent scheduling of data streams from each channel. It also utilizes hash feature dimensionality reduction to compress high-dimensional temperature feature vectors into a low-dimensional space, significantly reducing memory usage without losing key thermal drift information, thus eliminating the risk of memory overflow and concurrent processing bottlenecks caused by high-frequency sampling at the root.

[0061] Specifically, the principle of this invention is as follows: The reason why the solution of this invention can solve the above-mentioned technical problems is that, on the one hand, the fluid thermodynamic gradient descent fusion algorithm transforms the partial differential equation of thermal fluid convection into an energy dissipation optimization problem in the thermal resistance network topology, and accurately calculates the thermal drift compensation coefficient of each measurement channel through iterative convergence, thus establishing a quantitative relationship between temperature distribution and drift from the physical mechanism level; on the other hand, the thermal field topology compensation model uses a graph convolutional neural network to structurally model the heat conduction path between the nodes of the thermal element, so that the thermal drift prediction value output by the model simultaneously includes local heat accumulation characteristics and global thermal field distribution information, rather than the isolated temperature value of a single sensing point; in addition, the bidirectional long short-term memory network can capture the bidirectional hysteresis characteristics of the temperature gradient on the time axis, making up for the response lag defect of the static model under dynamic thermal disturbance; the incremental learning mechanism enables the model to update local parameters only for newly added deviation samples during long-term operation, avoiding the computational consumption of full retraining, while retaining historical compensation knowledge, so that the entire compensation method can remain effective under device aging conditions.

[0062] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0063] The specific implementation of step S01 is as follows: Multiple thermistors are deployed inside the marine multi-parameter precision measuring instrument, and each thermistor outputs a multi-channel temperature data stream using a high-frequency sampling method. A sliding window streaming processing architecture is used to segment, buffer, and batch process the data stream, with a fixed-length time window. Data from each channel is segmented, and an asynchronous coroutine mechanism is used to achieve concurrent scheduling of each channel, avoiding thread blocking. Assume there are a total of... The thermistor element, the first A thermistor at time The sampling temperature is The unit is ℃. Within a sliding window, the first... Channel collected Each sampling point forms a temperature vector. The unit is ℃. For the first... Channel temperature feature vector Perform a hash mapping to obtain a low-dimensional feature vector. The mapping formula is expressed as follows:

[0064] ;

[0065] In the formula For the first The channel's low-dimensional temperature feature vector is dimensionless and has a dimension of . ; for The dimensional hash projection matrix is ​​dimensionless, and its elements are generated by a random hash function. ; The reference temperature amplitude, in °C, is used for normalization. The low-dimensional eigenvectors of each thermistor are arranged row-wise to construct the internal temperature distribution matrix. for A dimensional matrix, whose dimensional matrix is ​​the first dimensional matrix. Behavior It is dimensionless, and its formula is as follows:

[0066]

[0067] The specific implementation of step S02 is as follows: Extract the partial differential equation of thermal fluid convection in the inner cavity of the marine multi-parameter precision measuring instrument, and transform it into a multidimensional temperature field state transition matrix, as expressed in the following formula:

[0068] ;

[0069] In the formula For the first The temperature of each thermistor element is expressed in °C. This is a reference temperature, in °C. The time step is expressed in seconds (s). For reference time, the unit is seconds; The thermal resistance network topology adjacency coefficient is dimensionless and is obtained by calibration of the internal thermal resistance network topology of the instrument. For the first A thermistor at time Temperature, in °C; For the first Local heat source power at the node, in W; Reference heat source power, unit is W; For the first Local heat capacity at nodes, in units of ; For reference heat capacity, the unit is... Using the real-time temperature readings of each thermistor in the internal temperature distribution matrix as input variables, a fluid thermodynamic gradient descent fusion algorithm is employed in the thermal resistance network topology to find the extreme points of the energy dissipation objective function. The formula for the energy dissipation objective function is as follows:

[0070] ;

[0071] In the formula Let be the objective function for energy dissipation, which is dimensionless. This is a reference loss value, dimensionless, and defaults to 1. This is the set of parameters to be optimized in the thermal resistance network topology, including... and wait; The first state transition matrix is ​​calculated based on the formula. Node predicted temperature, in °C; This represents the total number of time steps involved in the computation. The gradient descent parameter update formula is expressed as follows:

[0072] ;

[0073] In the formula For the first The parameter vector at the next iteration is dimensionless. The reference parameter scale is dimensionless and defaults to 1. For the objective function with respect to The gradient is dimensionless. The learning rate is a dimensionless, dimensionless gradient scale, defaulting to 1. The formula for adjusting the learning rate is as follows:

[0074] ;

[0075] In the formula For the first The learning rate for each iteration is dimensionless. The initial learning rate is dimensionless. The learning rate is dimensionless and serves as a reference. The attenuation factor is dimensionless. For the first During the nth iteration Local heat capacity parameters at nodes, in units of ; For reference heat capacity, the unit is... ,and The two have the same meaning and are consistent in the formula. This indicates that here The normalized benchmark, empirical value, is specifically used in the learning rate formula. The values ​​are consistent. After multiple rounds of iterative convergence, the thermal drift compensation coefficients for each measurement channel are output. , dimensionless.

[0076] The specific implementation of step S03 is as follows: The internal temperature distribution matrix... The measurement parameters collected from each measurement channel are input into the thermal field topology compensation model. The thermal field topology compensation model adopts a graph convolutional neural network structure, treating the thermal element as a graph node and the heat conduction path as the graph edge. Information aggregation and message passing iteration are performed between graph nodes through an adjacency matrix. Skip connections are used to prevent feature oversmoothing, and the predicted thermal drift values ​​of each measurement channel are output. The units should be consistent with the units of the corresponding measurement channel parameters. Channel Index Temperature, salinity, pressure, dissolved oxygen The deviation between the predicted thermal drift value and the current measurement parameters of each measurement channel is calculated. The formulas for calculating the mean deviation and the rate of change of deviation are expressed as follows:

[0077] ;

[0078] ;

[0079] In the formula For the first The average deviation within the current window of the channel, with the same unit as the corresponding channel; For reference deviation, the unit should be consistent with the corresponding channel; This represents the number of samples within the statistical window; For the first in the window Each corrected measurement data point has the same unit as the corresponding channel. For the corresponding number Each thermal drift prediction value is in the same unit as the corresponding channel. For the first Channel deviation change rate, in corresponding channel units ; The reference deviation rate is expressed in channel units. ; The time interval between adjacent statistical windows is expressed in seconds. The deviation is input into the adaptive adjustment function, and the formula for the comprehensive adjustment factor is as follows:

[0080] ;

[0081] In the formula For the first Channel-wide regulatory factor, dimensionless; For the first The channel confidence index, dimensionless, is calculated from the normalized activation variance of the output layer of the thermal field topology compensation model. For reference confidence level, dimensionless; , , is the weighting coefficient, dimensionless. When When, a large step weight update strategy is adopted; when When, a standard step-size weight update strategy is adopted; when When, a small-step weight update strategy is adopted; when At this time, pause the weight update and keep the weight parameters of the thermal field topology compensation model unchanged.

[0082] The specific implementation of step S04 is as follows: a bidirectional long short-term memory network is used to... The temperature time series data is used for bidirectional modeling (forward and backward), simultaneously capturing the dynamic hysteresis characteristics during the temperature gradient rise and fall processes, and outputting feedforward compensation values. The units are consistent with the units of the corresponding measurement channels. An adaptive Kalman filter is introduced. Transient thermal noise filtering is performed using adaptive Kalman filtering, which dynamically adjusts the filter gain based on the real-time estimation of the measurement noise covariance, in addition to standard Kalman filtering. This filters transient thermal noise introduced by drastic water temperature changes, yielding the filtered feedforward compensation value. The units should be consistent with the units of the corresponding measurement channels. Compared with thermal drift prediction Fusion, measuring parameters of each measurement channel Perform online thermal drift correction, and then measure the data after correction. The calculation formula is expressed as follows:

[0083] ;

[0084] In the formula For the first The current measurement parameters for the channel are displayed in units consistent with the corresponding channel. For the first The reference range for the channel should be consistent with the unit of the corresponding channel. The units for the corrected measurement data are consistent with those for the corresponding channels.

[0085] The specific implementation of step S05 is as follows: During the continuous operation of the marine multi-parameter precision measuring instrument, the dynamic offset of the thermal drift baseline of each measuring channel is monitored. The calculation formula for the dynamic offset of the thermal drift baseline is expressed as follows:

[0086] ;

[0087] In the formula For the first Channel at time The dynamic offset of the thermal drift baseline, with units consistent with the corresponding channel; For a moment The mean of the corrected measurement data within a nearby long-term window, with units consistent with the corresponding channel; The baseline update time interval is in seconds. Input incremental learning mechanism, only utilizing The corresponding new samples update the local weight parameters of the thermal topology compensation model without reconstructing all weight parameters. This reduces the computational cost of retraining while suppressing the forgetting of historical knowledge, thus maintaining the long-term effectiveness of the thermal topology compensation model.

[0088] The specific implementation of step S06 is as follows: During the laboratory calibration stage, multiple sets of temperature gradient excitations are applied to the marine multi-parameter precision measuring instrument, and the thermal drift amplitude of each measurement channel within the typical operating temperature range is recorded. The mean square error of the calibration data and the corrected measurement data under multiple sets of temperature gradient excitation conditions is statistically analyzed. The mean square error formula for consistency evaluation is expressed as follows:

[0089] ;

[0090] In the formula For the first The mean square error between the measured data and the calibration data after channel correction, expressed in squares of the corresponding channel units; This represents the number of temperature gradient excitation groups. For the first Channel 1 The units of the corrected measurement data under the group conditions are consistent with the corresponding channels. To correspond to the true value of thermal drift calibration, the units are consistent with the corresponding channels. Is it below the preset threshold? As a basis for determining whether the local heat capacity parameter in a thermal resistance network topology needs to be adjusted. A dimensionless, pre-defined consistency assessment threshold is set based on the instrument's measurement accuracy requirements. If the threshold is exceeded, the local heat capacity parameter is adjusted according to the assessment results. And the adjusted The feedback is sent to the fluid thermodynamic gradient descent fusion algorithm in step S02 to update the thermal drift compensation coefficient and achieve calibration closed loop.

[0091] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0092] To illustrate the technical solution of this invention, technicians set up a test environment and placed a marine multi-parameter precision measuring instrument integrating four measurement channels of temperature, salinity, pressure, and dissolved oxygen in a programmable temperature control box to simulate the typical temperature change scenarios experienced by the instrument during marine profile observation. The real-time temperature readings of each thermal element inside the instrument and the measurement parameters of each measurement channel were recorded throughout the process.

[0093] Eight negative temperature coefficient thermistors are deployed inside the instrument, installed in four areas: near the main control board processor, on the analog signal conditioning circuit board, on the sensor mounting base, and on the sealed cavity wall. Two thermistors are deployed in each area to improve spatial resolution. The sampling frequency of each thermistor is set to 20Hz, the sliding window length is set to 128 sampling points, the sliding step size is set to 32 sampling points, the feature dimension after hash feature dimensionality reduction is 30% of the original dimension, and the dimension of the internal temperature distribution matrix is ​​8×128.

[0094] The test consisted of two phases. The first phase was the laboratory calibration phase, in which the instrument was placed in a temperature-controlled chamber, and the temperature was increased from 5℃ to 35℃ at a rate of 1℃ / min, and then decreased to 5℃ at a rate of 2℃ / min, repeated 3 times. The thermal drift amplitude of each measurement channel during the heating and cooling process was recorded as calibration data, as shown in Table 1.

[0095] Table 1. Calibration data of thermal drift amplitude for each measurement channel

[0096]

[0097] After the first stage of calibration, temperature-drift sample pairs were constructed using the calibration data to train the thermal field topology compensation model. The thermal field topology compensation model employs a three-layer graph convolutional layer, with each layer followed by a residual skip connection. The loss function includes a principal term of mean squared error and a constraint term of the heat conduction law, with the constraint term weighted at 0.1. Mini-batch gradient descent is used for training, with a batch size of 32, iterating until the change in the loss function is less than [a certain value]. Training is stopped when the time is right. The initial learning rate of the fluid thermodynamic gradient descent fusion algorithm is 0.01, the decay factor is 0.5, and the thermal drift compensation coefficients for each measurement channel are shown in Table 2.

[0098] Table 2 Thermal drift compensation coefficients for each measurement channel

[0099]

[0100] The second stage is the dynamic compensation verification stage. The instrument is rapidly immersed from a 35℃ environment into 5℃ water at a rate of 3℃ / min to simulate the dramatic temperature changes experienced by the instrument during deep-sea profiling. The bidirectional long short-term memory network receives the current internal temperature distribution matrix as input. The forward and reverse channels each model the temperature time-series data using 64 hidden units, outputting feedforward compensation values. The initial value of the measurement noise covariance of the adaptive Kalman filter is set to 0.01, and it is adjusted in real time as the rate of temperature change increases. The adaptive adjustment function calculates a comprehensive adjustment factor based on the mean deviation, the rate of change of deviation, and the confidence index, such as... Figure 3 As shown, the comprehensive adjustment factor reaches a peak during the drastic change in water temperature, triggering a large step weight update strategy. The thermal field topology compensation model completes adaptive weight adjustment within about 90 seconds, and the drift amplitude of the measured data is significantly reduced after correction.

[0101] After 72 hours of continuous operation, a dynamic shift in the thermal drift baseline of the dissolved oxygen channel was detected, with a dynamic shift of 0.04 μmol / L, exceeding the preset threshold of 0.03 μmol / L. This triggered the incremental learning mechanism, which only locally updated the weight parameters of the output layer of the thermal field topology compensation model. The update step size was 0.1 times the baseline step size. After the update, the dynamic shift of the thermal drift baseline fell back below the preset threshold, and the long-term effectiveness of the thermal field topology compensation model was maintained.

[0102] The consistency evaluation results show that the mean square error of the corrected measurement data and the calibration data under all temperature gradient excitation conditions is less than 0.05, which meets the preset threshold requirements. There is no need to adjust the local thermal capacity parameters in the thermal resistance network topology, and the closed-loop calibration process does not trigger the process of resolving the thermal drift compensation coefficient.

[0103] Compared to traditional static calibration and compensation methods, this invention explicitly models the physical conduction topology of the thermal field as the input graph structure of a graph convolutional neural network. This allows the compensation model to perceive the spatial correlation between heat sources transmitted through heat conduction paths, rather than relying solely on isolated corrections based on single-point temperature readings. This results in stronger spatial awareness in multi-channel coupled thermal drift scenarios. By embedding the constraints of the heat conduction law into the loss function, prior physical knowledge constrains the search direction in the parameter space during training. This ensures that the model converges to a physically reasonable parameter region even with a limited sample size, fundamentally eliminating the overfitting risk of purely data-driven methods in sparse regions of the thermal field.

[0104] like Figure 2 As shown, the temperature time-series curves of each thermistor during the heating and cooling stages of the test are presented, intuitively demonstrating the spatial distribution differences and dynamic hysteresis characteristics of the internal thermal field. Figure 3 The figure shows the curve of the overall adjustment factor changing over time and the corresponding trigger interval of the weight update strategy. Figure 4 As shown, the curves compare the drift amplitude of each measurement channel before and after compensation, demonstrating the compensation effect of the present invention in dynamic thermal disturbance scenarios.

[0105] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0106] Table 3. Variable Explanation Table (Part 1)

[0107]

[0108] Table 4. Variable Explanation Table (Part Two)

[0109]

[0110] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for compensating for internal thermal drift in a marine multi-parameter precision measuring instrument, characterized in that, Includes the following steps: Multiple thermistors are deployed inside the marine multi-parameter precision measuring instrument. A sliding window streaming processing architecture combined with an asynchronous coroutine mechanism is used to schedule the multi-channel high-frequency temperature sampling data streams output by each thermistor. Hash feature dimensionality reduction is used to remove redundant features from the multi-channel high-frequency temperature sampling data streams to construct an internal temperature distribution matrix. The partial differential equation of thermal fluid convection in the inner cavity of the marine multi-parameter precision measuring instrument is extracted and transformed into a multidimensional temperature field state transition matrix. The real-time temperature readings of each thermosensitive element in the internal temperature distribution matrix are used as input variables. The extreme point of the energy dissipation objective function is found in the thermal resistance network topology using the fluid thermodynamic gradient descent fusion algorithm. The thermal drift compensation coefficient of each measurement channel is output. The internal temperature distribution matrix and the measurement parameters collected by each measurement channel are input into the thermal field topology compensation model, and the thermal field topology compensation model outputs the thermal drift prediction value of each measurement channel. The deviation between the predicted thermal drift value and the current measurement parameters of each measurement channel is calculated. The deviation is input into the adaptive adjustment function, and the weight parameters of the thermal field topology compensation model are dynamically adjusted according to the interval of the comprehensive adjustment factor output by the adaptive adjustment function. A bidirectional long short-term memory network is used to process the temperature time series data in the internal temperature distribution matrix, capture the dynamic hysteresis characteristics of the temperature gradient, and output the feedforward compensation value. An adaptive Kalman filter is introduced to filter the transient thermal noise of the feedforward compensation value. The filtered feedforward compensation value is fused with the thermal drift prediction value to perform online thermal drift correction on the measurement parameters of each measurement channel and output the corrected measurement data. During the continuous operation of the marine multi-parameter precision measuring instrument, the dynamic offset of the thermal drift baseline of each measuring channel is monitored. The dynamic offset is input into the incremental learning mechanism, and the thermal field topology compensation model is adjusted online only for the dynamic offset to maintain the long-term effectiveness of the thermal field topology compensation model. During the laboratory calibration phase, multiple temperature gradient excitations were applied to the marine multi-parameter precision measuring instrument, and the thermal drift amplitude of each measurement channel within the typical operating temperature range was recorded. The consistency between the calibration data and the corrected measurement data was evaluated, and the local thermal capacity parameters in the thermal resistance network topology were adjusted based on the consistency evaluation results. The adjusted local thermal capacity parameters were then fed back to the fluid thermodynamic gradient descent fusion algorithm.

2. The method for internal thermal drift compensation of a marine multi-parameter precision measuring instrument according to claim 1, characterized in that, The sliding window streaming architecture specifically refers to segmenting and batch processing multi-channel high-frequency temperature sampling data streams using a fixed-length time window, and using an asynchronous coroutine mechanism to achieve concurrent scheduling of data streams from each channel, thus avoiding thread blocking.

3. The method for compensating for internal thermal drift in a marine multi-parameter precision measuring instrument according to claim 2, characterized in that, The hash feature dimensionality reduction specifically refers to performing hash mapping on multidimensional temperature feature vectors, compressing high-dimensional features into a low-dimensional space, retaining the main thermal drift information, and eliminating redundant temperature features.

4. The method for compensating for internal thermal drift in a marine multi-parameter precision measuring instrument according to claim 3, characterized in that, The multidimensional temperature field state transition matrix describes the transmission relationship of the internal cavity temperature field between time steps. The temperature change of a thermistor over time is determined by the adjacency coefficient of the thermal resistance network topology, the power of the local heat source, and the local heat capacity. Each parameter is normalized to its corresponding reference value before being used in the calculation.

5. The method for compensating for internal thermal drift in a marine multi-parameter precision measuring instrument according to claim 4, characterized in that, The fluid thermodynamic gradient descent fusion algorithm specifically refers to using the real-time temperature readings of each thermistor as the state variable and energy dissipation as the objective function in the thermal resistance network topology. In each iteration, the learning rate step size is dynamically adjusted according to the local heat capacity parameters. After multiple rounds of convergence, the thermal drift compensation coefficients of each measurement channel are output.

6. The method for internal thermal drift compensation of a marine multi-parameter precision measuring instrument according to claim 5, characterized in that, The thermal field topology compensation model adopts a graph convolutional neural network structure, treating the thermal element as a graph node and the heat conduction path as the graph edge. It performs information aggregation and message passing iteration between graph nodes through an adjacency matrix, uses skip connections to prevent feature oversmoothing, and outputs the thermal drift prediction value of each measurement channel.

7. The method for internal thermal drift compensation of a marine multi-parameter precision measuring instrument according to claim 6, characterized in that, The establishment of the training dataset for the thermal field topology compensation model specifically includes: applying different temperature gradient excitations to the marine multi-parameter precision measuring instrument, synchronously recording the real-time temperature readings of each thermosensitive element, the measurement parameters of each measurement channel and the corresponding calibration thermal drift true values, and constructing temperature-drift sample pairs under multiple operating conditions.

8. The method for compensating for internal thermal drift in a marine multi-parameter precision measuring instrument according to claim 7, characterized in that, The training of the thermal field topology compensation model specifically uses the mean square error between the predicted thermal drift value and the calibrated true thermal drift value as the main term of the loss function. The constraint of the heat conduction law is embedded into the loss function in the form of a physical information neural network. The weight parameters of the thermal field topology compensation model are iteratively optimized using the mini-batch gradient descent method until the loss function converges.

9. The method for compensating for internal thermal drift in a marine multi-parameter precision measuring instrument according to claim 8, characterized in that, The adaptive adjustment function calculates the comprehensive adjustment factor using the mean deviation, the rate of change of deviation, and the confidence index as inputs. When the comprehensive adjustment factor is not less than 1.5, a large step weight update strategy is adopted; when the comprehensive adjustment factor is in the interval [0.8, 1.5), a standard step weight update strategy is adopted; when the comprehensive adjustment factor is in the interval [0.3, 0.8), a small step weight update strategy is adopted; and when the comprehensive adjustment factor is less than 0.3, weight update is paused.

10. The method for internal thermal drift compensation of a marine multi-parameter precision measuring instrument according to claim 9, characterized in that, The confidence index is specifically a quantitative indicator of the stability of the thermal drift prediction output of the thermal topology compensation model under the current input conditions, and is calculated from the normalized activation variance of the output layer of the thermal topology compensation model.

Images