Marine multi-parameter precision measuring instrument

Abstract

The application provides a kind of marine multi-parameter precision measuring instrument, belongs to marine measurement technical field, the application is integrated in the outer wall of sealed pressure cabin body platinum resistance temperature sensor, inductive conductivity sensor, silicon piezoresistive pressure sensor, dissolved oxygen optical sensor and pH electrode sensor, multi-parameter collaborative compensation and field reconstruction module are built-in in control chip, the modified pressure value is obtained by executing temperature and pressure cross-sensitivity compensation in turn, the abnormal probability score vector and space temperature and salinity interpolation field are output by temperature and salinity profile vector input artificial intelligence model, the attention temperature parameter of encoder is dynamically adjusted according to profile comprehensive anomaly index, finally, the multi-parameter profile data after quality control is written into data storage and communication module, solve the problem that multi-parameter marine profiler cannot complete high-precision salinity calculation conforming to TEOS-10 standard in real time on embedded low-power processor.

Description

Technical Field

[0001] This invention belongs to the field of marine measurement technology, and more specifically, relates to a marine multi-parameter precision measuring instrument. Background Technology

[0002] Multi-parameter oceanographic profiling instruments are core equipment for acquiring vertical distribution information of ocean elements such as temperature, conductivity, pressure, dissolved oxygen, and pH. They are widely used in autonomous observation platforms such as Argo buoys, underwater gliders, and moored buoys. In existing technologies, salinity is not directly measured but is calculated from the three elements of temperature, conductivity, and pressure based on the Thermodynamic Equations of the International System of Equations (TEOS-10). The TEOS-10 polynomial equations have high order and many terms, requiring extensive floating-point multiplications for direct expansion. On embedded oceanographic instruments equipped with low-power microcontrollers, the computing power of the floating-point units is limited, preventing the completion of full-precision TEOS-10 salinity calculations within each sampling cycle. Existing instruments typically use reduced-order approximation formulas or offline post-processing to avoid the real-time computational burden. The former sacrifices accuracy, while the latter cannot output salinity data in real time at the instrument, thus restricting the online quality control and real-time transmission capabilities of multi-parameter profiling data. In other words, existing technologies present a technical problem where multi-parameter oceanographic profiling instruments on embedded low-power processors cannot perform high-precision salinity calculations in real time. Summary of the Invention

[0003] In view of this, the present invention provides a marine multi-parameter precision measuring instrument, which can solve the technical problem that existing multi-parameter marine profilers cannot complete high-precision salinity calculation in real time on embedded low-power processors.

[0004] This invention is implemented as follows: It provides a marine multi-parameter precision measuring instrument, including a platinum resistance temperature sensor, an inductive conductivity sensor, a silicon piezoresistive pressure sensor, a dissolved oxygen optical sensor, a pH electrode sensor, a sample injection pump, a sealed pressure-resistant chamber, a lithium battery pack, a data storage and communication module, and a control chip. The sealed pressure-resistant chamber is a cylindrical metal shell closed at both ends. The control chip, lithium battery pack, and data storage and communication module are all fixedly installed inside the sealed pressure-resistant chamber. Sensor mounting bases are provided on the outer wall of the sealed pressure-resistant chamber. The device and pH electrode sensor are fixedly mounted on the sensor mounting base and are in direct contact with seawater. The injection pump is connected to the injection port of the inductive conductivity sensor through the injection pipe. The injection pump is fixedly mounted on the outer wall of the sealed pressure-resistant chamber and drives seawater to flow through the inductive conductivity sensor at a constant flow rate. The platinum resistance temperature sensor, inductive conductivity sensor, silicon piezoresistive pressure sensor, dissolved oxygen optical sensor, pH electrode sensor, and injection pump are all electrically connected to the control chip through waterproof cables passing through the sealed wiring holes of the sealed pressure-resistant chamber. The lithium battery pack is connected to the control chip, platinum resistance temperature sensor, inductive conductivity sensor, and silicon piezoresistive pressure sensor through the power management circuit. The pressure sensor, dissolved oxygen optical sensor, pH electrode sensor, and injection pump are electrically connected and provide power to these components. The data storage and communication module is electrically connected to the control chip, storing multi-parameter profile data output by the control chip and transmitting this data to an external platform via underwater acoustic or surface wireless communication. The platinum resistance temperature sensor employs a four-wire, dual-redundant platinum resistance structure; both the first and second platinum resistance signal lines are electrically connected to the high-precision analog-to-digital converter circuit within the control chip. The pressure signal output of the silicon piezoresistive pressure sensor is electrically connected to the control chip via a signal conditioning circuit. The piezoresistive pressure sensor incorporates a high-precision NTC thermistor. The temperature signal output terminal of the high-precision NTC thermistor is electrically connected to the control chip through a signal conditioning circuit. The control chip is equipped with a multi-parameter collaborative compensation and field reconstruction module. This module performs time alignment compensation, temperature and pressure cross-sensitivity compensation, platinum resistance drift self-calibration, real-time salinity calculation, and multi-platform sparse observation field reconstruction on the output signals of the platinum resistance temperature sensor, inductive conductivity sensor, silicon piezoresistive pressure sensor, dissolved oxygen optical sensor, and pH electrode sensor. It outputs multi-parameter profile data after quality control and writes it to the data storage and communication module.

[0005] Specifically, the time alignment compensation performed by the multi-parameter collaborative compensation and field reconstruction module refers to modeling the dynamic responses of the inductive conductivity sensor and the dissolved oxygen optical sensor as linear time-invariant systems, constructing a first inverse filter and a second inverse filter using the Wiener filtering method, performing a convolution operation between the original conductivity value and the first inverse filter to output a time-aligned conductivity value, and performing a convolution operation between the fluorescence quenching phase angle and the second inverse filter to output a time-aligned fluorescence quenching phase angle.

[0006] The dynamic transfer function parameters on which the construction of the first inverse filter and the second inverse filter are based are pre-stored in the non-volatile memory of the control chip and are dynamically called according to the current descent rate. The current descent rate is calculated by the ratio of the difference between the correction pressure value and the sampling time interval.

[0007] The temperature and pressure cross-sensitivity compensation based on the temperature and pressure dual-parameter cross-compensation coefficient matrix is ​​obtained by performing polynomial least squares fitting on the full-condition matrix calibration data in the pressure and temperature dual-axis calibration chamber. It is pre-stored in the non-volatile memory of the control chip and is called to perform multiplication and addition operations in the compensation calculation without the need for floating-point calculation unit participation.

[0008] The multi-parameter collaborative compensation and field reconstruction module includes the following steps: Simultaneously acquiring the resistance values ​​of the first and second platinum resistance thermometers, the original conductivity value, the original pressure voltage value, the sensor body temperature value, the fluorescence quenching phase angle, and the electrode potential at a sampling rate of 1000 Hz, forming an original multi-parameter time series; based on the sensor body temperature value, calling the temperature-pressure dual-parameter cross-compensation coefficient matrix to perform temperature-pressure cross-sensitivity compensation calculation on the original pressure voltage value, outputting a corrected pressure value, and calculating the current descent rate based on the ratio of the difference between the corrected pressure values ​​at consecutive sampling times to the sampling time interval; based on the current descent rate, calling the dynamic transfer function parameters of each sensor to perform dewinding on the original conductivity value channel and the fluorescence quenching phase angle channel respectively. The time alignment compensation operation outputs a time-aligned conductivity value and a time-aligned fluorescence quenching phase angle. These values ​​are then aligned on the depth axis to the sampling times corresponding to the resistance values ​​of the first and second platinum resistance thermometers. A differential resistance signal is extracted from the first and second platinum resistance values. Linear regression is performed on this differential signal over multiple consecutive sampling periods to obtain the differential drift trend slope. The absolute value of the differential drift trend slope is compared to a drift threshold. When the absolute value of the differential drift trend slope exceeds the drift threshold, a self-calibration program is triggered, controlling the chip to drive the reference liquid circulation in the sealed cavity. A loop is formed around the platinum resistance temperature sensor, collecting the resistance values ​​of the first and second channels as correction benchmarks. The zero-point compensation coefficient of the platinum resistance sensor is updated, and zero-point correction is performed on the first channel resistance value based on the updated zero-point compensation coefficient, outputting the corrected temperature value. The corrected temperature value, time-aligned conductivity value, and corrected pressure value are input into the TEOS-10 real-time salinity calculation unit based on Horner expansion and Q-format fixed-pointing, outputting the real-time salinity value. The corrected temperature value, real-time salinity value, and corrected pressure value are arranged in depth order to form a complete temperature-salinity-depth profile vector, which, along with geographic location information and seasonal coding, is input into the ocean profile quality assessment and field gravity. The ocean profile quality assessment and field reconstruction model outputs anomaly probability score vectors for each depth layer and a spatial temperature-salinity interpolation field obtained by sparse recovery of the graph signal. Based on the anomaly probability values ​​of each depth layer in the anomaly probability score vectors and the corresponding temperature-salinity gradient amplitudes, the comprehensive anomaly index of the profile is calculated. The comprehensive anomaly index of the profile is substituted into the profile quality adjustment function. The attention temperature parameter of the Transformer encoder in the ocean profile quality assessment and field reconstruction model is adjusted according to the range of the output value of the profile quality adjustment function. Finally, the multi-parameter profile data after quality control and the spatial temperature-salinity interpolation field are written into the data storage and communication module.

[0009] The TEOS-10 real-time salinity calculation unit based on Horner expansion and Q-format fixed-pointing reorganizes the calculation order of the multivariate polynomials involving corrected temperature values, time-aligned conductivity values, and corrected pressure values ​​in the TEOS-10 international thermodynamic equations using the Horner nested multiplication rule. This ensures that each additional polynomial order requires only one multiplication and one addition, reducing the number of multiplication operations from being proportional to the square of the polynomial order to being linearly proportional to the polynomial order.

[0010] Specifically, the Q-format fixed-point format scales the corrected temperature value, time-aligned conductivity value, and corrected pressure value to Q16 or Q24 fixed-point format according to their ranges. The intermediate calculation results are executed using 32-bit integer multiplication and addition instructions, eliminating the need for floating-point arithmetic units throughout the process.

[0011] The profile comprehensive anomaly index is obtained by normalizing the weighted summation of the anomaly probability values ​​of each depth layer in the anomaly probability scoring vector with the corresponding temperature and salinity gradient magnitude as the weight. The greater the temperature and salinity gradient magnitude, the greater the contribution of the depth layer to the profile comprehensive anomaly index.

[0012] The profile quality adjustment function is a piecewise linear function with the profile comprehensive anomaly index as the independent variable and the attention temperature adjustment coefficient as the dependent variable. The product of the attention temperature adjustment coefficient and the attention temperature parameter of the Transformer encoder constitutes the updated attention temperature parameter. When the profile comprehensive anomaly index is in the range [0, 0.3), the attention temperature adjustment coefficient takes a first preset value. When the profile comprehensive anomaly index is in the range [0.3, 0.7), the attention temperature adjustment coefficient takes a second preset value. When the profile comprehensive anomaly index is not less than 0.7, the attention temperature adjustment coefficient takes a third preset value. The first preset value is less than the second preset value, and the second preset value is less than the third preset value.

[0013] The marine profile quality assessment and field reconstruction model consists of two parts: a profile anomaly detection and quality control subnetwork and a multi-platform data fusion and field reconstruction subnetwork. The profile anomaly detection and quality control subnetwork takes the complete temperature, salinity, and depth profile vector, geographic location information, and seasonal coding as inputs. It inputs the complete temperature, salinity, and depth profile vector into a multi-layer Transformer encoder with the depth dimension as the sequence length. The multi-layer Transformer encoder outputs a full profile context representation vector. The full profile context representation vector is followed by an anomaly scoring multi-layer sensing head. The anomaly scoring multi-layer sensing head outputs an anomaly probability score vector.

[0014] The multi-platform data fusion and field reconstruction sub-network constructs a graph structure from the temperature and salinity measurements, node geographic coordinates, and timestamps of external multi-observation platform nodes. It uses the GraphSAGE algorithm to aggregate the temperature and salinity features of neighboring nodes, realizes feature propagation between sparse observation nodes, and outputs a spatial temperature and salinity interpolation field. The graph structure calculates edge weights based on the spatial distance and time difference between nodes and supports dynamic addition and subtraction of nodes.

[0015] The graph signal sparse recovery abstracts the geographical coordinates and timestamps of multiple external observation platforms into graph nodes. It calculates the edge weights between nodes using a Gaussian kernel function that is jointly weighted by the geographical distance and depth difference between nodes, constructs the graph Laplacian matrix of the irregular spatial graph, performs eigenvalue decomposition on the graph Laplacian matrix, extracts the low-frequency graph spectral basis vectors corresponding to smaller eigenvalues, applies a low-pass filter constraint in the graph frequency domain, and iteratively solves the problem using the alternating direction multiplier method to output the spatial temperature-salinity interpolation field.

[0016] The training of the ocean profile quality assessment and field reconstruction model includes: pre-training a multilayer Transformer encoder on an unsupervised pre-training dataset with masked profile reconstruction as the pre-training target; fine-tuning anomaly scoring multilayer sensing head on a supervised fine-tuning dataset, using binary cross-entropy as the loss function; and training the graph aggregation parameters of the GraphSAGE algorithm using graph structure training samples as input and high-resolution reanalysis temperature and salinity field data as supervision labels, employing the mean square error loss function.

[0017] The unsupervised pre-training dataset is constructed by extracting temperature, salinity, and depth profile data marked with official quality control from the global Argo Project historical database and stratifying them by geographical location and season. The supervised fine-tuning dataset is constructed by selecting profile data containing salt spikes, density inversion, and sensor saturation anomaly markers from the global Argo Project historical database and profile data confirmed as high quality by manual review, and stratifying them by anomaly type.

[0018] The differential drift trend slope is obtained by performing linear regression on the differential sequence formed by the difference between the resistance values ​​of the first and second platinum resistance thermometers over multiple consecutive sampling periods. When the absolute value of the differential drift trend slope exceeds the drift threshold, it is determined that the platinum resistance temperature sensor has experienced irreversible baseline drift, triggering a self-calibration program. The self-calibration program updates the zero-point compensation coefficient of the platinum resistance thermometer under the known reference liquid temperature condition.

[0019] The profile anomaly detection and quality control subnetwork and the multi-platform data fusion and field reconstruction subnetwork are jointly fine-tuned. The joint fine-tuning ensures that the anomaly probability scoring vector and the spatial temperature-salinity interpolation field are mutually constrained in terms of physical consistency. The loss function of the joint fine-tuning is composed of a weighted sum of binary cross-entropy loss and mean square error loss.

[0020] This invention employs a TEOS-10 real-time salinity calculation unit based on Horner nested expansion and Q-scheme fixed-point formatting. This reduces the computational complexity of TEOS-10 polynomials from being proportional to the square of the order to being linearly proportional to the order. Simultaneously, all intermediate operations are converted into 32-bit integer multiplication and addition instructions, completely eliminating the reliance on floating-point units. In traditional direct expansion, each additional polynomial order requires an increase in the number of multiplications proportional to the number of existing terms. Horner nested expansion, however, fixes the computational load for each additional order at one multiplication and one addition, reducing the total number of multiplications from the quadratic to the linear order. Q-scheme fixed-point formatting scales physical quantities such as temperature, conductivity, and pressure to the Q16 or Q24 integer domain according to their range. Intermediate calculations are performed entirely using integer multiplication and addition instructions, allowing the low-power microcontroller to complete salinity calculations within each sampling interval without needing to activate the floating-point coprocessor. In summary, this invention solves the technical problem mentioned in the background art that multi-parameter ocean profilers cannot perform high-precision salinity calculations conforming to the TEOS-10 standard in real time on embedded low-power processors. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the measuring instrument of the present invention.

[0022] Figure 2 This is a flowchart of the method of the present invention.

[0023] Figure 3 A comparison of the relationship between temperature and conductivity profile depth before and after time alignment compensation.

[0024] Figure 4 The diagram shows the anomaly probability scores and temperature-salinity gradient amplitude distributions for each depth layer. Detailed Implementation

[0025] 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.

[0026] like Figure 1The diagram shown is a structural schematic of a marine multi-parameter precision measuring instrument provided by the present invention. The instrument includes: a platinum resistance temperature sensor, an inductive conductivity sensor, a silicon piezoresistive pressure sensor, a dissolved oxygen optical sensor, a pH electrode sensor, a sample pump, a sealed pressure-resistant chamber, a lithium battery pack, a data storage and communication module, and a control chip. The sealed pressure-resistant chamber is a cylindrical metal shell closed at both ends. The control chip, the lithium battery pack, and the data storage and communication module are all fixedly installed inside the sealed pressure-resistant chamber. A sensor mounting base is provided on the outer wall of the sealed pressure-resistant chamber. The platinum resistance temperature sensor, the inductive conductivity sensor, the silicon piezoresistive pressure sensor, the silicon piezoresistive pressure sensor, the dissolved oxygen optical sensor, the pH electrode sensor, the sample pump, a sealed pressure-resistant chamber, a lithium battery pack, a data storage and communication module, and a control chip. The conductivity sensor, the silicon piezoresistive pressure sensor, the dissolved oxygen optical sensor, and the pH electrode sensor are respectively fixedly mounted on the sensor mounting base and are in direct contact with seawater. The sampling pump is connected to the inlet of the inductive conductivity sensor through a sampling pipeline. The sampling pump is fixedly mounted on the outer wall of the sealed pressure-resistant chamber and drives seawater to flow through the inductive conductivity sensor at a constant flow rate. The platinum resistance temperature sensor, the inductive conductivity sensor, the silicon piezoresistive pressure sensor, the dissolved oxygen optical sensor, the pH electrode sensor, and the sampling pump are all passed through the seal of the sealed pressure-resistant chamber by waterproof cables. The wiring hole is electrically connected to the control chip; the lithium battery pack is electrically connected to the control chip, the platinum resistance temperature sensor, the inductive conductivity sensor, the silicon piezoresistive pressure sensor, the dissolved oxygen optical sensor, the pH electrode sensor, and the sample injection pump through a power management circuit, and provides operating power to the above components; the data storage and communication module is electrically connected to the control chip, used to store the multi-parameter profile data output by the control chip and transmit the multi-parameter profile data to an external platform through underwater acoustic communication or surface wireless communication; the platinum resistance temperature sensor adopts a four-wire dual-redundant platinum resistance structure, and the platinum resistance temperature... The first and second platinum resistance signal lines of the pressure sensor are both electrically connected to the high-precision analog-to-digital converter circuit within the control chip, for synchronously providing the first and second platinum resistance resistance values ​​to the control chip; the pressure signal output terminal of the silicon piezoresistive pressure sensor is electrically connected to the control chip through a signal conditioning circuit, for providing the control chip with the raw pressure voltage value; the silicon piezoresistive pressure sensor incorporates a high-precision NTC thermistor, and the temperature signal output terminal of the high-precision NTC thermistor is electrically connected to the control chip through the signal conditioning circuit, for providing the control chip with the sensor body temperature value;The control chip incorporates a multi-parameter collaborative compensation and field reconstruction module. This module performs time alignment compensation, temperature and pressure cross-sensitivity compensation, platinum resistance temperature sensor drift self-calibration, real-time salinity calculation, and multi-platform sparse observation field reconstruction on the first and second platinum resistance resistance values ​​output by the platinum resistance temperature sensor, the raw conductivity value output by the inductive conductivity sensor, the raw pressure voltage value and sensor body temperature value output by the silicon piezoresistive pressure sensor, the fluorescence quenching phase angle output by the dissolved oxygen optical sensor, and the electrode potential output by the pH electrode sensor. Finally, it outputs quality-controlled multi-parameter profile data and writes it to the data storage and communication module.

[0027] like Figure 2 As shown, the multi-parameter collaborative compensation and field reconstruction module is used to perform the following steps:

[0028] S01, the control chip synchronously acquires the first and second platinum resistance values ​​output by the platinum resistance temperature sensor, the raw conductivity value output by the inductive conductivity sensor, the raw pressure voltage value and sensor body temperature value output by the silicon piezoresistive pressure sensor, the fluorescence quenching phase angle output by the dissolved oxygen optical sensor, and the electrode potential output by the pH electrode sensor at a sampling rate of 1000 Hz, forming a raw multi-parameter time series;

[0029] S02. Based on the temperature value of the sensor body, call the temperature and pressure dual-parameter cross-compensation coefficient matrix pre-stored in the non-volatile memory of the control chip, perform temperature and pressure cross-sensitivity compensation operation on the original pressure voltage value, output the corrected pressure value, and calculate the current descent rate based on the ratio of the difference between the corrected pressure value and the sampling time interval.

[0030] S03. Based on the current descent rate, call the dynamic transfer function parameters of each sensor pre-stored in the non-volatile memory of the control chip, and perform deconvolution time alignment compensation operation on the conductivity raw value channel and fluorescence quenching phase angle channel in the original multi-parameter time series respectively, output the time-aligned conductivity value and the time-aligned fluorescence quenching phase angle, and align the time-aligned conductivity value and the time-aligned fluorescence quenching phase angle on the depth axis to the sampling time corresponding to the first platinum resistance value and the second platinum resistance value, thereby eliminating the depth profile misalignment caused by sensor response time mismatch;

[0031] S04. Extract the differential resistance signal from the first and second platinum resistance values. Perform linear regression on the differential resistance signal over multiple consecutive sampling periods to obtain the differential drift trend slope. Compare the absolute value of the differential drift trend slope with a drift threshold pre-stored in the non-volatile memory of the control chip. When the absolute value of the differential drift trend slope exceeds the drift threshold, trigger the self-calibration program. The control chip drives the reference liquid in the sealed cavity to circulate around the platinum resistance temperature sensor. Under the condition of known reference liquid temperature, collect the first and second platinum resistance values ​​as correction benchmarks, update the platinum resistance zero-point compensation coefficient, perform zero-point correction on the first platinum resistance value based on the updated platinum resistance zero-point compensation coefficient, and output the corrected temperature value.

[0032] S05. Input the corrected temperature value, the time-aligned conductivity value, and the corrected pressure value into the TEOS-10 real-time salinity calculation unit based on Horner expansion and Q-format fixed-pointing, and output the real-time salinity value.

[0033] S06. The corrected temperature value, the real-time salinity value, and the corrected pressure value are arranged in depth order to form a complete temperature-salinity-depth profile vector, which is then input into the ocean profile quality assessment and field reconstruction model along with the geographic location information and seasonal code. The ocean profile quality assessment and field reconstruction model outputs the anomaly probability score vector corresponding to each depth layer and the spatial temperature-salinity interpolation field obtained by sparse recovery of the graph signal.

[0034] S07. Based on the anomaly probability values ​​of each depth layer in the anomaly probability scoring vector and the corresponding temperature and salinity gradient amplitudes of each depth layer, calculate the comprehensive profile anomaly index. Substitute the comprehensive profile anomaly index into the profile quality adjustment function, and adjust the attention temperature parameter of the Transformer encoder in the ocean profile quality assessment and field reconstruction model according to the range of the output value of the profile quality adjustment function: when the comprehensive profile anomaly index is in the range [0, 0.3), lower the attention temperature parameter to concentrate the multi-head attention weight distribution of the Transformer encoder on the anomaly depth layer, enhancing the response sensitivity to local anomalies; when the comprehensive profile anomaly index is in the range [0.3, 0.7), keep the attention temperature parameter at a moderate level to maintain balanced attention across the entire profile context; when the comprehensive profile anomaly index is ≥0.7, raise the attention temperature parameter to make the multi-head attention weight distribution more uniform, reducing over-response to local noise and thus reducing the false alarm rate; finally, write the quality-controlled multi-parameter profile data and the spatial temperature and salinity interpolation field into the data storage and communication module.

[0035] Specifically, the temperature and pressure dual-parameter cross-compensation coefficient matrix refers to a coefficient matrix obtained by performing polynomial least squares fitting on the full-condition matrix calibration data in a pressure and temperature dual-axis calibration chamber, with the sensor body temperature value and the original pressure voltage value as independent variables and the pressure zero-point offset and sensitivity correction coefficient as dependent variables. The temperature and pressure dual-parameter cross-compensation coefficient matrix is ​​pre-stored in the non-volatile memory of the control chip and is called in step S02 to perform multiplication and addition operations without the participation of a floating-point calculation unit, thereby realizing real-time temperature and pressure compensation on the control chip.

[0036] Specifically, the deconvolution time alignment compensation operation refers to modeling the dynamic responses of the inductive conductivity sensor and the dissolved oxygen optical sensor as linear time-invariant systems, respectively, constructing a first inverse filter corresponding to the inductive conductivity sensor channel and a second inverse filter corresponding to the dissolved oxygen optical sensor channel using the Wiener filtering method, performing a convolution operation between the original conductivity value and the first inverse filter to output the time-aligned conductivity value, and performing a convolution operation between the fluorescence quenching phase angle and the second inverse filter to output the time-aligned fluorescence quenching phase angle, so that the time-aligned conductivity value and the time-aligned fluorescence quenching phase angle are aligned on the depth axis to the equivalent instantaneous sampling time of the platinum resistance temperature sensor, thereby eliminating the systematic depth misalignment caused by the flow rate lag of the injection tube and the electrochemical response delay of the optical sensor.

[0037] Specifically, the differential drift trend slope refers to the regression slope obtained by performing linear regression on the differential sequence formed by the difference between the resistance values ​​of the first platinum resistance and the second platinum resistance over multiple consecutive sampling periods. When the absolute value of the differential drift trend slope exceeds the drift threshold, it is determined that the platinum resistance temperature sensor has experienced irreversible baseline drift, triggering the self-calibration program.

[0038] Specifically, the TEOS-10 real-time salinity calculation unit based on Horner expansion and Q-format fixed-point format is implemented as follows: the multivariate polynomials involving the corrected temperature value, the time-aligned conductivity value, and the corrected pressure value in the TEOS-10 international thermodynamic equations are reorganized using the Horner nested multiplication rule, so that each additional polynomial order only requires one multiplication and one addition, reducing the number of multiplication operations from being proportional to the square of the polynomial order to being linearly proportional to the polynomial order; at the same time, the physical quantities of the corrected temperature value, the time-aligned conductivity value, and the corrected pressure value are scaled to Q16 or Q24 fixed-point format according to their ranges, and the intermediate calculation results are executed with 32-bit integer multiplication and addition instructions, without the need for a floating-point arithmetic unit, thereby completing the calculation of the real-time salinity value on the control chip within each sampling cycle. The technical benefits of the TEOS-10 real-time salinity calculation unit based on Horner expansion and Q-format fixed-point format are as follows: Traditional TEOS-10 polynomial direct expansion requires a large number of floating-point multiplications, which is time-consuming and power-intensive on embedded processors, making real-time salinity calculation impossible at high sampling rates. By adopting Horner nested expansion, the polynomial computation complexity is reduced from being proportional to the square of the order to being linearly proportional to the order. Combined with Q-format fixed-point format, all intermediate operations are transformed into integer multiplication and addition, enabling the low-power microcontroller to complete salinity calculation in each sampling interval. This ensures that the salinity calculation accuracy is consistent with the TEOS-10 standard while significantly reducing computational power consumption, extending the long-term deployment endurance under lithium battery power supply conditions, and providing an engineering-feasible implementation path for high-frequency real-time salinity output.

[0039] The specific implementation of the ocean sparse observation field reconstruction algorithm based on graph signal processing is as follows: The geographical coordinates and timestamps of multiple external observation platforms are abstracted as graph nodes. The edge weights between nodes are calculated using a Gaussian kernel function weighted by the geographical distance and depth difference between nodes, constructing a graph Laplacian matrix of an irregular spatial graph. Eigenvalue decomposition is performed on the graph Laplacian matrix to extract low-frequency graph basis vectors corresponding to smaller eigenvalues. These low-frequency graph basis vectors correspond to the large-scale spatial structure of the ocean temperature and salinity field. A low-pass filter constraint is applied in the graph frequency domain, transforming the field reconstruction problem of sparse temperature and salinity observation nodes into a convex optimization problem under low-frequency graph basis constraints. The ADMM alternating direction multiplier method is used for iterative solution, outputting the spatial temperature and salinity interpolation field. The technical advantage of the ocean sparse observation field reconstruction algorithm based on graph signal processing for the entire scheme is that traditional optimal interpolation methods rely on a pre-assumed fixed correlation length scale, leading to decreased reconstruction accuracy when observation nodes dynamically change or are sparsely distributed. The graph signal processing method directly encodes the physical spatial relationships between observation nodes into a graph structure. The low-frequency graph basis vectors adaptively capture the large-scale ocean temperature and salinity spatial structure under the current node distribution. The ADMM alternating direction multiplier method iterative solution has fast convergence in sparse node scenarios and the results satisfy the physical smoothness constraint. At the same time, the graph structure supports the dynamic addition and removal of nodes. When a new observation node enters or leaves the network, only the graph Laplacian matrix needs to be locally updated without recalibrating the global parameters, so that the entire system has the ability to adaptively reconstruct the field in multi-platform collaborative observation scenarios.

[0040] Specifically, the profile comprehensive anomaly index refers to the scalar value obtained by normalizing the weighted summation of the anomaly probability values ​​of each depth layer in the anomaly probability scoring vector with the corresponding temperature and salinity gradient amplitude as the weight. The greater the temperature and salinity gradient amplitude of the depth layer, the greater its contribution to the profile comprehensive anomaly index. As a result, the anomalies in physically sensitive areas such as thermoclines have a more significant impact on the profile comprehensive anomaly index than those in non-gradient areas.

[0041] Specifically, the profile quality adjustment function refers to a piecewise linear function with the profile comprehensive anomaly index as the independent variable and the attention temperature adjustment coefficient as the dependent variable. The product of the attention temperature adjustment coefficient and the attention temperature parameter of the Transformer encoder constitutes the updated attention temperature parameter. When the profile comprehensive anomaly index belongs to [0, 0.3), the attention temperature adjustment coefficient takes a first preset value; when the profile comprehensive anomaly index belongs to [0.3, 0.7), the attention temperature adjustment coefficient takes a second preset value; and when the profile comprehensive anomaly index ≥ 0.7, the attention temperature adjustment coefficient takes a third preset value. The first preset value is less than the second preset value, and the second preset value is less than the third preset value.

[0042] The specific structure of the ocean profile quality assessment and field reconstruction model is as follows: the ocean profile quality assessment and field reconstruction model consists of two parts: a profile anomaly detection and quality control subnetwork and a multi-platform data fusion and field reconstruction subnetwork. The profile anomaly detection and quality control subnetwork takes the complete temperature, salinity, and depth profile vector, the geographical location information, and the seasonal code as inputs, and inputs the complete temperature, salinity, and depth profile vector into a multi-layer Transformer encoder with the depth dimension as the sequence length. The multi-layer Transformer encoder adopts a multi-head self-attention mechanism, which simultaneously refers to the physical state of other depth layers in the entire profile when assessing anomalies at a single depth layer, and establishes the temperature structure at the thermocline and the subsurface layer. The long-range physical coupling relationship between salinity structures is defined. The multi-layer Transformer encoder outputs a full-profile context representation vector, which is then followed by an anomaly scoring multi-layer sensing head. The anomaly scoring multi-layer sensing head outputs the anomaly probability scoring vector. The multi-platform data fusion and field reconstruction sub-network constructs a graph structure from the temperature and salinity measurements, node geographic coordinates, and timestamps of external multi-observation platform nodes. The GraphSAGE algorithm is used to aggregate the temperature and salinity features of neighboring nodes, enabling feature propagation between sparse observation nodes and outputting the spatial temperature and salinity interpolation field. The graph structure jointly calculates the edge weights based on the spatial distance and time difference between nodes and supports dynamic addition and removal of nodes.

[0043] The steps for establishing the training dataset for the ocean profile quality assessment and field reconstruction model specifically include: extracting officially quality-controlled temperature, salinity, and depth (TDM) profile data from the global Argo program historical database, stratifying and sampling by geographical location and season to construct an unsupervised pre-training dataset; selecting profile data containing salt spikes, density inversion, and sensor saturation anomaly markers, as well as manually reviewed and confirmed high-quality profile data from the global Argo program historical database, stratifying by anomaly type to construct a supervised fine-tuning dataset; and extracting TDM measurements and corresponding node geographic coordinates and timestamps from multiple observation platforms within the same time window from the multi-platform synchronous observation cruise database to construct graph structure training samples, using high-resolution reanalysis TDM field data as the supervised label for the graph reconstruction task.

[0044] The specific steps of training the ocean profile quality assessment and field reconstruction model include: pre-training the multilayer Transformer encoder on the unsupervised pre-training dataset with masked profile reconstruction as the pre-training target, so that the multilayer Transformer encoder learns the depth-direction physical covariance structure of the global climate profile; fine-tuning the anomaly scoring multilayer sensing head with the supervised fine-tuning dataset, using binary cross-entropy as the loss function and anomaly labels of each depth layer as supervision signals; training the graph aggregation parameters of the GraphSAGE algorithm with the graph structure training samples as input and the high-resolution reanalysis temperature-salinity field data as supervision labels, using the mean square error loss function; and finally, jointly fine-tuning the profile anomaly detection and quality control subnetwork and the multi-platform data fusion and field reconstruction subnetwork to ensure that the anomaly probability scoring vector and the spatial temperature-salinity interpolation field are mutually constrained in terms of physical consistency.

[0045] The technical benefits of the ocean profile quality assessment and field reconstruction model to the overall scheme are as follows: The multi-head self-attention mechanism of the multi-layer Transformer encoder enables the ocean profile quality assessment and field reconstruction model to simultaneously refer to the physical state of other depth layers in the entire profile when assessing anomalies at a single depth layer. This establishes long-range physical coupling relationships between the thermocline, halocline, and subsurface structure, avoiding missed detections and false alarms caused by the traditional layer-by-layer threshold judgment method that ignores the overall profile structure. The GraphSAGE algorithm explicitly encodes the spatial correlation of sparse observations from multiple platforms into a graph structure, aggregates neighborhood information, and outputs a continuous spatial temperature and salinity interpolation field. This compensates for the insufficient spatial representativeness of single-instrument profile observations, allowing the multi-parameter profile data output by a single measuring instrument to be integrated into a larger-scale ocean field reconstruction, thereby enhancing the comprehensive utilization value of the observation data. The joint training of the profile anomaly detection and quality control subnetwork and the multi-platform data fusion and field reconstruction subnetwork enables the anomaly probability scoring vector and the spatial temperature and salinity interpolation field to mutually verify each other at the physical constraint level, further improving the reliability of the entire system's output data.

[0046] The specific implementation of step S01 is as follows: The control chip synchronously triggers the acquisition of signals from each sensor at a sampling rate of 1000 Hz. The first and second platinum resistance values ​​of the platinum resistance temperature sensor are synchronously sampled by a high-precision analog-to-digital conversion circuit. The original conductivity value output by the inductive conductivity sensor, the original pressure voltage value and sensor body temperature value output by the silicon piezoresistive pressure sensor, the fluorescence quenching phase angle output by the dissolved oxygen optical sensor, and the electrode potential output by the pH electrode sensor are all acquired at the same sampling trigger time. The sampling data of each channel are aligned with timestamps and stored in a circular buffer to form an original multi-parameter time series, providing a unified time reference for the original data input of each subsequent compensation step.

[0047] The specific implementation of step S02 is as follows: The temperature and pressure dual-parameter cross-compensation coefficient matrix is ​​obtained before leaving the factory by performing polynomial least squares fitting on the full-condition matrix calibration data through a pressure and temperature dual-axis calibration chamber. The coefficient matrix is ​​pre-stored in the non-volatile memory of the control chip. In each sampling cycle, the control chip reads the sensor body temperature value as the row index and the original pressure voltage value as the input. Based on the coefficient matrix, it performs multiplication and addition operations to calculate the pressure zero-point offset and sensitivity correction coefficient, completes temperature and pressure cross-sensitivity compensation for the original pressure voltage value, and outputs the corrected pressure value. The entire operation only calls integer multiplication and addition instructions, without the need for a floating-point unit, meeting the real-time processing constraints of the low-power microcontroller. The reference range for the corrected pressure value is 0–6000 dbar, and the reference value for the drift threshold is set to 0.005% of the sensor's full scale. In a continuous sampling sequence, the difference between the corrected pressure values ​​at adjacent sampling times is divided by the sampling time interval to obtain the current rate of decrease, in units of... This is used for the dynamic calling of subsequent deconvolution filter parameters.

[0048] The specific implementation of step S03 is as follows: The dynamic responses of the inductive conductivity sensor and the dissolved oxygen optical sensor are modeled as linear time-invariant systems, respectively. Their time constants are measured under different descent rates during factory calibration, and the corresponding dynamic transfer function parameters are pre-stored in the non-volatile memory of the control chip. Based on the current descent rate obtained in step S02, the corresponding dynamic transfer function parameters are retrieved from the non-volatile memory, and the first inverse filter corresponding to the inductive conductivity sensor channel and the second inverse filter corresponding to the dissolved oxygen optical sensor channel are constructed using the Wiener filtering method. The original conductivity value is convolved with the first inverse filter to output a time-aligned conductivity value; the fluorescence quenching phase angle is convolved with the second inverse filter to output a time-aligned fluorescence quenching phase angle. The deconvolution operation aligns the two signals on the depth axis to the equivalent instantaneous sampling time of the platinum resistance temperature sensor, eliminating the systematic depth misalignment caused by the flow rate lag of the injection tube and the electrochemical response delay of the optical sensor. The typical time constant reference value for an inductive conductivity sensor is approximately 0.05–0.1 s, and the typical time constant reference value for a dissolved oxygen optical sensor is approximately 2–5 s. The reference value for the regularization parameter of the Wiener filter is set to... It is used to suppress high-frequency noise amplification during the deconvolution process.

[0049] The specific implementation of step S04 is as follows: The difference between the first and second synchronously acquired platinum resistance values ​​is calculated point by point to form a dual-channel resistance difference sequence. Linear regression is performed on the difference sequence over N consecutive sampling periods, where N is a reference value of 500–2000 sampling points, corresponding to a duration of 0.5–2 s. The regression slope is the slope of the differential drift trend. The absolute value of the differential drift trend slope is compared with a pre-stored drift threshold, with a reference value of 0.01 mΩ / s. When the threshold is exceeded, it is determined that the platinum resistance temperature sensor has experienced irreversible baseline drift, triggering a self-calibration procedure. The self-calibration program drives the reference liquid in the sealed cavity to circulate around the platinum resistance temperature sensor. The reference liquid temperature is read in real time by an independent high-precision thermometer. Under the condition of known reference liquid temperature, the resistance values ​​of the first and second platinum resistance sensors are collected. The difference between the collected value and the theoretical resistance value is used as the zero-point compensation amount, and the zero-point compensation coefficient of the platinum resistance sensor is updated. Based on the updated zero-point compensation coefficient, the resistance value of the first platinum resistance sensor is zero-point corrected, and the corrected temperature value is output. The range of the corrected temperature value is -2 to 35℃, and the resolution is better than 0.001℃.

[0050] The specific implementation of step S05 is as follows: The corrected temperature value, time-aligned conductivity value, and corrected pressure value are input into the TEOS-10 real-time salinity calculation unit based on Horner expansion and Q-format fixed-point scaling. The TEOS-10 salinity equation system consists of multivariate high-order polynomials. Direct expansion requires a multiplication number proportional to the square of the order, resulting in a large computational load. The Horner nested multiplication rule is used to reorganize the evaluation order of each polynomial, so that each additional polynomial order requires only one multiplication and one addition, reducing the number of multiplications to a linear proportion to the order. The corrected temperature value is scaled to Q16 fixed-point scaling within the range of -2 to 40℃, and the time-aligned conductivity value is scaled to the range of 0 to 7℃. The scale is adjusted to Q16 fixed points, and the pressure value is adjusted to Q24 fixed points according to the range of 0 to 6000 dbar. All intermediate operations are executed with 32-bit integer multiplication and addition instructions, without the need for a floating-point unit. Real-time salinity values ​​are output within each 1000 Hz sampling interval. The accuracy of the real-time salinity values ​​is consistent with the TEOS-10 standard, and the range reference is 30 to 38 PSU.

[0051] The specific implementation of step S06 is as follows: The corrected temperature value, real-time salinity value, and corrected pressure value are arranged in order of depth from shallow to deep to form a complete temperature, salinity, and depth profile vector. The reference value for the number of profile depth layers is 500–2000 layers, and the reference value for the step size is 1–5 m. The complete temperature, salinity, and depth profile vector, geographical location information, and seasonal coding are input into the ocean profile quality assessment and field reconstruction model. The profile anomaly detection and quality control subnetwork, with depth as the sequence length, inputs the complete temperature, salinity, and depth profile vector layer by layer into a multi-layer Transformer encoder. The multi-head self-attention mechanism simultaneously considers the remaining depth layers of the entire profile when evaluating any single depth layer, establishing a long-range physical coupling relationship between the thermocline, halocline, and subsurface structure, and outputting a full-profile context representation vector. This vector is then connected to an anomaly scoring multi-layer sensing head, outputting the anomaly probability value for each depth layer to form an anomaly probability scoring vector. The multi-platform data fusion and field reconstruction sub-network synchronously receives temperature and salinity measurements, node geographic coordinates, and timestamps from multiple external observation platform nodes. It constructs a graph Laplacian matrix by calculating the edge weights between nodes using a Gaussian kernel function, extracts low-frequency spectrum basis vectors, and iteratively solves the convex optimization problem using the alternating direction multiplier method to output the spatial temperature and salinity interpolation field.

[0052] The specific implementation of step S07 is as follows: The anomaly probability values ​​of each depth layer in the anomaly probability scoring vector are weighted, summed, and normalized using the corresponding temperature-salinity gradient amplitude as weights to obtain the profile comprehensive anomaly index, with a value range of [0, 1]. The profile comprehensive anomaly index is substituted into the profile quality adjustment function, which is a piecewise linear function: when the profile comprehensive anomaly index belongs to [0, 0.3), the attention temperature adjustment coefficient takes a first preset value, with a reference value of 0.5, to concentrate the multi-head attention weights of the Transformer encoder on the anomaly depth layer; when the profile comprehensive anomaly index belongs to [0.3, 0.7), the attention temperature adjustment coefficient takes a second preset value, with a reference value of 1.0, to maintain balanced attention across the entire profile; when the profile comprehensive anomaly index is not less than 0.7, the attention temperature adjustment coefficient takes a third preset value, with a reference value of 2.0, to make the multi-head attention weights more uniform, reduce excessive response to local noise, and reduce the false alarm rate. The product of the attention temperature adjustment coefficient and the current attention temperature parameter constitutes the updated attention temperature parameter. This updated parameter is applied to the Softmax normalization operation of the Transformer encoder, adjusting the concentration of the attention weight distribution. Finally, the quality-controlled multi-parameter profile data and the spatial temperature-salinity interpolation field are written into the data storage and communication module, completing the entire data processing workflow for this profile observation.

[0053] It should be noted that the present invention includes the following key technical concepts:

[0054] Firstly, it utilizes Horner expansion and Q-format fixed-point encoding for real-time TEOS-10 salinity calculation. Traditional direct expansion methods involve multiplication numbers proportional to the square of the polynomial order, which cannot meet the computational power requirements of high-frequency real-time computing on low-power microcontrollers. Horner nested expansion compresses the number of multiplications to a linear proportion to the order, while Q-format fixed-point encoding maps all floating-point operations to the integer multiplication-addition domain. The combined effect of these two methods allows salinity calculation to be completed within each sampling period, maintaining TEOS-10 accuracy while significantly reducing computational power consumption.

[0055] Secondly, dual-channel platinum resistance differential drift self-calibration. A single-channel platinum resistance sensor cannot distinguish between actual temperature changes and sensor baseline drift. The difference between dual-channel differential signals is theoretically zero when the physical temperature changes uniformly. Performing linear regression on the differential sequence can sensitively capture irreversible baseline drift. Combined with online self-calibration using the reference liquid in the sealed cavity, the zero-point compensation coefficient can be updated without water discharge, eliminating the accuracy degradation caused by long-term deployment.

[0056] Third, sparse observation field reconstruction based on graph signal processing across multiple platforms. Traditional optimal interpolation relies on a preset fixed correlation length scale, which reduces reconstruction accuracy when nodes change dynamically. Graph signal processing directly encodes the spatial relationships of observation nodes into a graph structure. The low-frequency graph basis vectors adaptively capture the large-scale temperature and salinity spatial structure under the current node distribution. When nodes are added or removed, only the graph Laplacian matrix needs to be updated locally. The alternating direction multiplier method converges quickly in sparse node scenarios and satisfies physical smoothness constraints.

[0057] The synergistic effect of the three technical approaches is as follows: Horner unrolling and Q-format fixed-pointing ensure high-quality salinity data output in real time on the embedded end, providing accurate node observation inputs for subsequent graph reconstruction subnetworks; differential drift self-calibration ensures long-term stability of the temperature measurement benchmark, guaranteeing the accuracy of profile measurements from the source, thereby improving the anomaly detection confidence of the Transformer encoder; graph signal processing field reconstruction integrates the vertical profile observations of a single instrument into the multi-platform spatial field, which in turn provides spatial context constraints for the Transformer encoder, enabling cross-validation of anomaly probability scores at the physical field level. The combined effect of these three technologies enables the entire system to achieve closed-loop processing from raw signal acquisition to multi-platform field reconstruction on a low-power embedded platform.

[0058] It should be noted that this invention also solves the following technical problem: In the long-term autonomous deployment of current marine multi-parameter profiling instruments, platinum resistance temperature sensors experience irreversible baseline drift due to the cumulative effects of material creep, electrochemical corrosion, and mechanical vibration. Traditional single-channel platinum resistance structures cannot distinguish between actual temperature changes and the sensor's own drift signal. The drift amount cannot be detected before the instrument leaves the water, resulting in compromised data quality during long-term deployment. Existing technologies suffer from the technical problem of not being able to detect and correct baseline drift of platinum resistance temperature sensors online under long-term deployment conditions. This invention employs a four-wire dual-redundant platinum resistance structure, extracting the drift trend slope from the differential resistance signals of the two channels. When the differential drift exceeds a threshold, it triggers online self-calibration of the reference liquid in the sealed cavity, completing the zero-point compensation coefficient update without requiring the instrument to leave the water, thus eliminating the impact of long-term drift on temperature measurement accuracy from a fundamental mechanism perspective.

[0059] Furthermore, during the high-speed descent of multi-sensor profiling instruments, the inductive conductivity sensor exhibits a flow rate lag in the injection tube, and the dissolved oxygen optical sensor suffers from an electrochemical response delay. This leads to inconsistent effective sampling depths for different parameters, resulting in a systematic misalignment on the depth axis. Existing technologies rely on fixed-time offset correction, which cannot adapt to dynamic changes in the descent rate. Existing technologies suffer from the technical problem of being unable to dynamically correct the systematic misalignment on the depth axis caused by response time mismatch in multi-sensor profiling data. This invention dynamically calls the corresponding dynamic transfer function parameters based on the real-time calculated current descent rate, and uses deconvolution time alignment compensation calculations to align the conductivity and dissolved oxygen data on the depth axis to the equivalent instantaneous sampling time of the platinum resistance temperature sensor, achieving adaptive depth alignment that changes with the descent rate.

[0060] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems is that the Horner nested multiplication rule reorganizes the evaluation order of multivariate polynomials from a mathematical structure perspective. It transforms the exponential operations that originally required independent calculation of each term into a chain structure of layer-by-layer multiplication and addition. This reduces the total number of multiplications required to calculate the same polynomial from being proportional to the square of the order to being linearly proportional to the order. This mathematical property is independent of the polynomial order, and therefore the benefits are particularly significant for high-order TEOS-10 equation systems. Q-format fixed-point representation solves the mapping problem between floating-point and integer operations at the hardware instruction set level. By scaling the real values ​​of each physical quantity to a fixed decimal point integer representation according to the range ratio, all intermediate operations can be completed by calling the microcontroller's native 32-bit integer multiplication and addition instructions. This preserves sufficient numerical dynamic range and precision while completely avoiding the overhead of calling the floating-point coprocessor. The synergistic effect of the two enables the low-power embedded processor to complete the full TEOS-10 salinity calculation within each 1000 Hz sampling period, fundamentally eliminating the contradiction between real-time performance and accuracy, and is logically feasible.

[0061] 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.

[0062] The specific implementation of step S01 is as follows: the control chip synchronously acquires the first channel platinum resistance value output by the platinum resistance temperature sensor at a sampling rate of 1000 Hz. The resistance value of the second platinum resistance The raw conductivity value output by the inductive conductivity sensor The original pressure voltage value output by the silicon piezoresistive pressure sensor and sensor body temperature value Fluorescence quenching phase angle output by dissolved oxygen optical sensor and the electrode potential output by the pH electrode sensor This forms the original multi-parameter time series, denoted as... ,in This is the sampling time sequence number.

[0063] The specific implementation of step S02 is as follows: the temperature and pressure dual-parameter cross-compensation coefficient matrix uses the sensor body temperature value as the basis. and pressure original voltage value With pressure zero-point offset and sensitivity correction coefficient as independent variables, the compensation calculation formula is obtained by performing polynomial least squares fitting on the full-condition matrix calibration data in the pressure and temperature biaxial calibration chamber. The compensation calculation formula is expressed as follows:

[0064] ;

[0065] In the formula, For dimensionless correction of pressure values; The elements of the temperature and pressure dual-parameter cross-compensation coefficient matrix are obtained through calibration experiments and fitting. The empirical value of the matrix order is... , Let the order of the polynomial in the temperature direction be . The order of the voltage-direction polynomial; This is a temperature normalization reference value, taken as the median value within the calibrated temperature range, in °C. This is a normalized reference value for pressure and voltage, in V. This is a normalized pressure reference value, in dbar. The dimension is dbar, making the entire right-hand side of the formula dimensionless. (Rate of descent) The formula is calculated as follows: It is the ratio of the corrected pressure difference between two consecutive sampling times to the sampling time interval.

[0066] ;

[0067] In the formula, The sampling time interval, s; As a reference value for the normalized descent rate, the empirical value is 0.1 m / s. Normalized to a dimensionless quantity, used for table lookup indexing of inverse filter coefficients.

[0068] The specific implementation of step S03 is as follows: the dynamic responses of the inductive conductivity sensor and the dissolved oxygen optical sensor are modeled as linear time-invariant systems, respectively, and the first inverse filter is constructed using the Wiener filtering method. Second inverse filter The formula is expressed as follows:

[0069] ;

[0070] ;

[0071] In the formula, Angular frequency, in rad / s; and The dynamic transfer functions of the conductivity sensor and the dissolved oxygen sensor are determined by the dynamic transfer function parameters of each sensor pre-stored in the non-volatile memory. and These are the power spectral densities of the conductivity signal and the dissolved oxygen signal, respectively. and These are the power spectral densities of the corresponding channel noise; all of the above power spectral densities were estimated by performing discrete Fourier transform on the output signal under static input of the sensor during the factory calibration stage and are pre-stored in non-volatile memory. and All are dimensionless frequency domain transfer function ratios. Time-aligned conductivity values. Time-aligned fluorescence quenching phase angle The output from the convolution operation is expressed by the following formula:

[0072] ;

[0073] ;

[0074] In the formula, and The first and second inverse filters are respectively at the current descent rate. The time-domain impulse response coefficients under the given conditions are obtained by analyzing the time-domain impulse response coefficients. and The time-domain impulse response was obtained by performing an inverse discrete Fourier transform, and multiple typical impulse responses were tested during the factory calibration phase. The values ​​are calculated separately and stored in non-volatile memory. At runtime, the values ​​are determined based on the current... Obtained by linear interpolation; For filter tap index; The coefficient is dimensionless. and Dimensions consistent, unit S / m; The coefficient is dimensionless. and The dimensions are consistent, and the unit is rad.

[0075] The specific implementation of step S04 is as follows: From and Extracting differential signals In continuous Within each sampling period Perform linear regression, difference drift trend slope The calculation formula is expressed as follows:

[0076] ;

[0077] In the formula, This is the number of sampling points participating in the regression; an empirical value is 1000. For the first The dual-path differential signal at time t, in Ω; The unit is Ω, representing the linear change in differential resistance within each sampling period. When Exceeding the pre-stored drift threshold When an irreversible baseline drift is detected in the platinum resistance temperature sensor, a self-calibration program is triggered. The control chip drives the reference liquid in the sealed cavity to circulate around the platinum resistance temperature sensor, at a known reference liquid temperature. Data collection under certain conditions and As a correction benchmark, update the zero-point compensation coefficient of the platinum resistance thermometer. The formula is expressed as follows:

[0078] ;

[0079] In the formula, The resistance value of the first platinum resistance thermometer collected during self-calibration timing is in Ω; for The standard theoretical resistance of a platinum resistance thermometer at a given temperature, in Ω, is calculated according to the industrial standard calibration formula for platinum resistance thermometers, as follows:

[0080] ;

[0081] In the formula, This is the standard resistance value of a platinum resistance thermometer at 0°C, for a Pt100 type sensor. Ω; The first-order temperature coefficient of platinum resistance thermometer has a standard value of ℃ ; The second temperature coefficient of platinum resistance thermometer has a standard value of [value missing]. ℃ ; The temperature normalization reference value is set to 1℃, making all terms in the formula dimensionless. The temperature of the known reference liquid is in °C. The unit is Ω. Based on the updated... right Perform zero-point correction; the corrected resistance value is Substitute it into the inverse function of the standard calibration formula for platinum resistance thermometers to solve for the corrected temperature value. , in °C.

[0082] The specific implementation of step S05 is as follows: The corrected temperature value is... Time-aligned conductivity value and dimensionless corrected pressure value Input is given to the International Thermodynamic Equations-10 real-time salinity calculation unit based on Horner expansion and Q-scheme fixed-pointing, to... , and (Dimensionless) is a dimensionless independent variable. The calculation order of the multivariate polynomial is reorganized using Horner's nested multiplication rule. An example formula for univariate directional expansion is expressed as follows:

[0083] ;

[0084] In the formula, These are the coefficients of the -10 polynomials of the International Thermodynamics Equations, with dimensions consistent with the salinity unit g / kg; This is a normalized reference value for temperature, in °C. The conductivity is a normalized reference value, and the empirical value is taken as the conductivity of standard seawater, which is 4.2914 S / m. The order is polynomial, with an empirical value of 5; the three-variable cross terms are also expanded layer by layer using Horner's nested method. Each physical quantity is scaled to its range. or Fixed-point number format This represents a fixed-point number format with a 16-digit decimal part. The number is a fixed-point number with a 24-bit decimal part. Intermediate results are executed using 32-bit integer multiplication and addition instructions, and the final output is the real-time salinity value. , in g / kg.

[0085] The specific implementation of step S06 is: to correct the temperature value Real-time salinity value and dimensionless corrected pressure value The complete temperature, salinity, and depth profile vector is formed in depth order, as expressed by the following formula:

[0086] ;

[0087] In the formula, For the complete temperature-salinity depth profile vector; , , The first Corrected temperature values, real-time salinity values, and dimensionless corrected pressure values ​​for the depth layer; This represents the number of layers in the profile depth. Including geographic location information and seasonal coding, the ocean profile quality assessment and field reconstruction model is used for graph signal sparsity recovery, with node edge weights. Calculated using the Gaussian kernel function, the formula is as follows:

[0088] ;

[0089] In the formula, For nodes With nodes The geographical distance between them, in km; For nodes With nodes The depth difference between them, in meters; This is a dimensionless bandwidth parameter representing the geographic distance direction. This is the bandwidth parameter in the depth direction, which is dimensionless. Normalized reference value for geographic distance, in km; This is a depth-normalized reference value, in meters (m). The edge weights are dimensionless and their values ​​range from 1 to 2. Graph Laplace matrix From the adjacency weight matrix The structure and formula are as follows:

[0090] ;

[0091] In the formula, For a degree matrix, its diagonal elements , For the reason The adjacency weight matrix formed; , , All A dimensionless square matrix This represents the total number of nodes in the graph. Perform eigenvalue decomposition , The eigenvector matrix, Given a diagonal matrix composed of corresponding eigenvalues, take the corresponding first... The eigenvectors of the smallest eigenvalues ​​form the low-frequency spectrum basis matrix. , The number of low-frequency components is empirically estimated to be 20% of the number of nodes. The formula for the field reconstruction optimization problem is as follows:

[0092] ,satisfy ;

[0093] In the formula, Let be the vector of the spatial temperature-salinity interpolation field to be determined; This is a vector of temperature and salinity measurements at sparse observation nodes; This is a normalized reference value for the thermo-salinity field, with dimensions equal to... Consistency makes all terms dimensionless; This is the regularization weight coefficient, dimensionless, with an empirical value of 0.01; For graph smoothing regularization, the interpolation field is constrained to satisfy physical smoothness in the graph structure; The low-frequency spectral coefficient vector has the same dimensions as... Consistent. The alternating direction multiplier method is used to iteratively solve the above convex optimization problem, introducing auxiliary variables. and dual variables The iterative formula is expressed as follows:

[0094] ;

[0095] ;

[0096] ;

[0097] In the formula, This represents the number of iterations. The Lagrange penalty parameter for the alternating direction multiplier method is dimensionless and has an empirical value of 1.0. The soft thresholding shrinkage operator is applied element-wise to the input vector. Operations, including The threshold value is soft and dimensionless. As an auxiliary variable, As dual variables, both are initialized to zero vectors; the convergence condition for iteration is two consecutive iterations. The difference between the updated L2 norm and The ratio is less than the convergence threshold, and the empirical value is .

[0098] The specific implementation method of step S07 is: profile comprehensive anomaly index The anomaly probability values ​​at each depth layer are weighted and normalized using the temperature and salinity gradient magnitude as the weights. The formula is as follows:

[0099] ;

[0100] In the formula, For the first The anomaly probability value of the deep layer has a range of values. Dimensionless; For the first The amplitude of the temperature-salinity gradient at depth, in °C / m or g / (kg·m); As a reference value for gradient normalization, the average gradient of the entire profile is taken, with dimensions equal to... Consistency, making Dimensionless; It is a dimensionless scalar, and its range of values ​​is 1. The attention temperature adjustment coefficient is based on the output of the profile quality adjustment function. With attention temperature parameter The product constitutes the updated attention temperature parameter The formula is expressed as follows:

[0101] ;

[0102] In the formula, Here, represents the initial value of the attention temperature parameter, and represents the scaling factor for the scaling dot product in the multi-head self-attention calculation of the converter encoder; empirically, is . ,in The key vector dimension of the attention head is initialized during model training and stored in the non-volatile memory of the control chip. For attention temperature regulation coefficient, dimensionless, when hour Take the first preset value ,when hour Take the second preset value ,when hour Take the third preset value ,in The empirical values ​​are 0.6, 1.0, and 1.5 respectively; For the updated attention temperature parameter, the dimensions are... Consistency; multi-parameter profile data after quality control and spatial temperature-salinity interpolation field It is also written into the data storage and communication module.

[0103] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0104] Technicians set up a test environment and selected a marine multi-parameter precision measuring instrument manufactured according to the present invention. The instrument was mounted on an underwater profiling buoy platform and a complete sinking profile observation was completed in an open sea area. The sinking depth was 1500 m and the average sinking rate was about 0.1 m / s. The instrument's operating temperature range covered 2 to 28℃. During the sinking process, each sensor continuously collected data synchronously at a sampling rate of 1000 Hz.

[0105] The instrument's hardware components include: a platinum resistance temperature sensor employing a four-wire, dual-redundant platinum resistance structure, with a measurement range of -2 to 35℃ and a resolution better than 0.001℃; and an inductive conductivity sensor with a measurement range of 0 to 7℃. The accuracy is better than 0.001. The silicon piezoresistive pressure sensor has a range of 0–2000 dbar and incorporates a high-precision NTC thermistor for body temperature compensation; the dissolved oxygen optical sensor is based on the fluorescence quenching principle and has a range of 0–500 μmol / L; the pH electrode sensor has a range of 6.5–9.0; the injection pump flow rate is set to a constant 100 mL / min; the sealed pressure-resistant chamber is an aluminum alloy cylindrical shell with a pressure resistance depth of 2000 m; the lithium battery pack has a capacity of 40 Wh; the control chip is a low-power 32-bit microcontroller with a main frequency of 200 MHz and no independent floating-point coprocessor; the data storage and communication module supports both underwater acoustic communication and Iridium satellite communication.

[0106] In step S01, the control chip synchronously triggers all sensor channels at 1000 Hz, and the original multi-parameter time series is stored in a 32 KB circular buffer with timestamp alignment for real-time access in subsequent compensation steps.

[0107] In step S02, the temperature and pressure dual-parameter cross-compensation coefficient matrix is ​​calibrated under full operating conditions in a pressure and temperature dual-axis calibration chamber before leaving the factory. The calibration matrix covers temperatures from -2 to 35℃ and pressures from 0 to 2000 dbar, with a total of 25×20 grid points, a polynomial order of 4, and a least-squares fitting root mean square residual better than 0.1 dbar. During online operation, the control chip performs coefficient matrix lookup and multiplication-addition operations using integer multiplication-addition instructions. A single temperature and pressure compensation operation takes approximately 2 μs. After correcting the pressure value output, the pressure difference between adjacent sampling times is divided by the sampling interval to obtain the current descent rate. In this test, the current descent rate remained stable between 0.09 and 0.11 m / s.

[0108] In step S03, the time constant of the inductive conductivity sensor is approximately 0.06 s under the conditions of this test, and the time constant of the dissolved oxygen optical sensor is approximately 3.2 s. The corresponding Wiener inverse filter regularization parameter is taken as... Based on the current descent rate of 0.10 m / s, the corresponding dynamic transfer function parameters are invoked to construct the first and second inverse filters, respectively. Deconvolution operations are performed on the original conductivity value and the fluorescence quenching phase angle, shifting the two signals forward by approximately 6 mm and 320 mm respectively along the depth axis, aligning them to the equivalent instantaneous sampling depth of the platinum resistance temperature sensor, thus eliminating depth misalignment. Figure 3 As shown, the depth correspondence between the conductivity profile and the temperature profile at the thermocline is significantly improved before and after time alignment, and the salinity spike artifact at the thermocline disappears.

[0109] In step S04, the control chip performs linear regression on the difference sequence formed by the difference between the resistance values ​​of the first and second platinum resistance thermometers over 1000 consecutive sampling points (i.e., a 1-second window). During the first 600 m depth of this test, the absolute value of the differential drift trend slope remained consistently below the drift threshold of 0.01 mΩ / s, indicating that the sensor was functioning normally. At a depth of approximately 800 m, the absolute value of the differential drift trend slope rose to 0.018 mΩ / s, exceeding the drift threshold and triggering a self-calibration program. The control chip then circulated the reference liquid in the sealed cavity around the platinum resistance temperature sensor. The reference liquid temperature, read as 8.52℃ by an independent reference thermometer, was used to acquire the resistance values ​​of both platinum resistance thermometers under known temperature conditions and update the zero-point compensation coefficient of the platinum resistance thermometers. The corrected temperature value and the deviation from the reference temperature converged to within 0.003℃, completing the online drift correction.

[0110] In step S05, the temperature value of 8.52℃ is corrected and scaled to Q16 fixed-point format, and the time-aligned conductivity value of 3.621 is used. Scaling to Q16 fixed-point format, correcting the pressure value to 802.3 dbar, and scaling to Q24 fixed-point format, the TEOS-10 polynomial is organized using Horner nested expansion. All intermediate operations are completed using 32-bit integer multiplication and addition instructions. A single salinity calculation takes approximately 48 μs, and the real-time salinity value output is 34.82 PSU. The deviation from the floating-point double-precision calculation result is less than 0.0003 PSU, and the accuracy meets the TEOS-10 standard requirements.

[0111] In step S06, a complete temperature, salinity, and depth (TDT) profile vector was constructed for this test profile at 1-m step sizes, totaling 1500 layers. This vector, along with geographic location information (marked with sea area codes) and seasonal coding (winter), was input into the oceanographic profile quality assessment and field reconstruction model. Simultaneously, TDT observation data from three surrounding Argo buoys and one moored buoy were received from four external nodes. The edge weights between nodes were calculated using a Gaussian kernel function to construct... The Graph Laplacian matrix is ​​used to extract the first three low-frequency spectrum basis vectors. The solution is obtained iteratively using the alternating direction multiplier method. The convergence criterion is that the root mean square change in the field reconstruction results of adjacent iteration steps is less than... The PSU converged after approximately 42 iterations in this test, outputting the spatial temperature-salinity interpolation field. Table 1 shows the anomaly probability scores for each depth layer, listing the anomaly probability values ​​and temperature-salinity gradient magnitudes for representative depth layers.

[0112] Table 1. Anomaly Probability Scores and Temperature-Salinity Gradient Amplitudes for Representative Depth Layers

[0113]

[0114] In step S07, based on the data in Table 1, the anomaly probability values ​​are weighted, summed, and normalized using the temperature and salinity gradient amplitudes at each depth layer as weights. The calculated comprehensive anomaly index for the profile is 0.38, falling within the range of [0.3, 0.7). The attention temperature adjustment coefficient is set to the second preset value of 1.0 to maintain balanced attention across the entire profile, keeping the Transformer encoder attention temperature parameter at a moderate level. Figure 4 As shown, the thermocline corresponds to a depth of 120 m. The anomaly probability value of 0.41 at this location reflects the sensitivity of the strong gradient structure above and below the thermocline to the local salinity structure. The Transformer encoder, through a multi-head self-attention mechanism, simultaneously references the physical states of the 50 m and 200 m depth layers when evaluating the 120 m depth layer, effectively establishing a long-range coupling relationship between the thermocline and the subsurface layer, and avoiding false alarms for this layer caused by isolated layer-by-layer threshold judgment methods. Finally, the complete 1500-layer multi-parameter profile data after quality control, along with the spatial temperature-salinity interpolation field, is written into the data storage and communication module and transmitted to an external platform via underwater acoustic communication.

[0115] Compared with traditional methods, this invention brings the following technological advancements: In real-time salinity calculation, traditional instruments cannot complete TEOS-10 real-time calculations on low-power embedded processors due to the burden of floating-point operations. However, Horner expansion and Q-format fixed-point transformation reduce computational complexity from both mathematical structure and instruction set perspectives, making it an engineering reality for low-power microcontrollers to complete TEOS-10 compliant salinity calculations in each sampling interval. In terms of long-term stability of temperature sensors, the dual-path differential structure and online self-calibration mechanism distinguish between real temperature changes and baseline drift from the signal detection principle, avoiding the limitations of traditional single-path structures that rely on periodic water discharge calibration. In terms of data quality assurance, the multi-head self-attention mechanism of the Transformer encoder upgrades anomaly detection from isolated layer-by-layer judgment to full-profile context awareness, effectively reducing missed detections and false alarms in physically sensitive areas such as thermoclines. In terms of multi-platform collaborative observation, graph signal processing field reconstruction integrates the vertical observation of a single instrument into a multi-node spatial field, compensating for the inherent deficiency of insufficient spatial representativeness of a single instrument. Furthermore, the graph structure supports dynamic addition and removal of nodes without the need to recalibrate global parameters. It should be noted that the user data involved in the embodiments of this application have all been authorized, acquired, processed, and transmitted in accordance with legal and regulatory requirements.

[0116] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0117] Table 2. Variable Explanation Table (Part 1)

[0118]

[0119] Table 3. Variable Explanation Table (Part Two)

[0120]

[0121] 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 marine multi-parameter precision measuring instrument, characterized in that, The system includes a platinum resistance temperature sensor, an inductive conductivity sensor, a silicon piezoresistive pressure sensor, a dissolved oxygen optical sensor, a pH electrode sensor, a sampling pump, a sealed pressure chamber, a lithium battery pack, a data storage and communication module, and a control chip. The sealed pressure chamber is a cylindrical metal shell closed at both ends. The control chip, lithium battery pack, and data storage and communication module are all fixedly installed inside the sealed pressure chamber. Sensor mounting bases are provided on the outer wall of the sealed pressure chamber. The platinum resistance temperature sensor, inductive conductivity sensor, silicon piezoresistive pressure sensor, dissolved oxygen optical sensor, and pH electrode sensor are fixedly installed on the sensor mounting bases and are in direct contact with seawater. The sampling pump is connected to the inlet of the inductive conductivity sensor through a sampling pipe. The sampling pump is fixedly installed on the outer wall of the sealed pressure chamber and drives seawater to flow through the inductive conductivity sensor at a constant flow rate. The platinum resistance temperature sensor, inductive conductivity sensor, silicon piezoresistive pressure sensor, and dissolved oxygen optical sensor are all present. The pH electrode sensor and injection pump are electrically connected to the control chip via waterproof cables passing through sealed wiring holes in the sealed pressure-resistant chamber. The data storage and communication module is electrically connected to the control chip, used to store multi-parameter profile data output by the control chip and transmit multi-parameter profile data to an external platform via underwater acoustic communication or surface wireless communication. The platinum resistance temperature sensor adopts a four-wire dual-redundant platinum resistance structure, and both the first and second platinum resistance signal lines of the platinum resistance temperature sensor are electrically connected to the high-precision analog-to-digital conversion circuit within the control chip. The pressure signal output terminal of the silicon piezoresistive pressure sensor is electrically connected to the control chip via a signal conditioning circuit. The silicon piezoresistive pressure sensor has a built-in high-precision NTC thermistor, and the temperature signal output terminal of the high-precision NTC thermistor is electrically connected to the control chip via a signal conditioning circuit. The control chip has a multi-parameter collaborative compensation and field reconstruction module, used to output multi-parameter profile data after quality control and write it to the data storage and communication module.

2. The marine multi-parameter precision measuring instrument according to claim 1, characterized in that, The time alignment compensation performed by the multi-parameter collaborative compensation and field reconstruction module specifically refers to modeling the dynamic responses of the inductive conductivity sensor and the dissolved oxygen optical sensor as linear time-invariant systems, constructing a first inverse filter and a second inverse filter using the Wiener filtering method, performing a convolution operation between the original conductivity value and the first inverse filter to output a time-aligned conductivity value, and performing a convolution operation between the fluorescence quenching phase angle and the second inverse filter to output a time-aligned fluorescence quenching phase angle.

3. The marine multi-parameter precision measuring instrument according to claim 2, characterized in that, The dynamic transfer function parameters on which the construction of the first inverse filter and the second inverse filter are based are pre-stored in the non-volatile memory of the control chip and are dynamically called according to the current descent rate. The current descent rate is calculated by the ratio of the difference between the correction pressure value and the sampling time interval.

4. The marine multi-parameter precision measuring instrument according to claim 3, characterized in that, The temperature and pressure cross-sensitivity compensation based on the temperature and pressure dual-parameter cross-compensation coefficient matrix is ​​obtained by performing polynomial least squares fitting on the full-condition matrix calibration data in the pressure and temperature dual-axis calibration chamber. It is pre-stored in the non-volatile memory of the control chip and is called to perform multiplication and addition operations in the compensation calculation without the need for floating-point calculation unit participation.

5. The marine multi-parameter precision measuring instrument according to claim 4, characterized in that, The multi-parameter collaborative compensation and field reconstruction module includes the following steps: Simultaneously acquiring the resistance values ​​of the first and second platinum resistance thermometers, the original conductivity value, the original pressure voltage value, the sensor body temperature value, the fluorescence quenching phase angle, and the electrode potential at a sampling rate of 1000 Hz, forming an original multi-parameter time series; based on the sensor body temperature value, calling the temperature-pressure dual-parameter cross-compensation coefficient matrix to perform temperature-pressure cross-sensitivity compensation calculation on the original pressure voltage value, outputting a corrected pressure value, and calculating the current descent rate based on the ratio of the difference between the corrected pressure values ​​at consecutive sampling times to the sampling time interval; based on the current descent rate, calling the dynamic transfer function parameters of each sensor to perform dewinding on the original conductivity value channel and the fluorescence quenching phase angle channel respectively. The time alignment compensation operation outputs a time-aligned conductivity value and a time-aligned fluorescence quenching phase angle. These values ​​are then aligned on the depth axis to the sampling times corresponding to the resistance values ​​of the first and second platinum resistance thermometers. A differential resistance signal is extracted from the first and second platinum resistance values. Linear regression is performed on this differential signal over multiple consecutive sampling periods to obtain the differential drift trend slope. The absolute value of the differential drift trend slope is compared to a drift threshold. When the absolute value of the differential drift trend slope exceeds the drift threshold, a self-calibration program is triggered, controlling the chip to drive the reference liquid circulation in the sealed cavity. A loop is formed around the platinum resistance temperature sensor, collecting the resistance values ​​of the first and second channels as correction benchmarks. The zero-point compensation coefficient of the platinum resistance sensor is updated, and zero-point correction is performed on the first channel resistance value based on the updated zero-point compensation coefficient, outputting the corrected temperature value. The corrected temperature value, time-aligned conductivity value, and corrected pressure value are input into the TEOS-10 real-time salinity calculation unit based on Horner expansion and Q-format fixed-pointing, outputting the real-time salinity value. The corrected temperature value, real-time salinity value, and corrected pressure value are arranged in depth order to form a complete temperature-salinity-depth profile vector, which, along with geographic location information and seasonal coding, is input into the ocean profile quality assessment and field gravity. The ocean profile quality assessment and field reconstruction model outputs anomaly probability score vectors for each depth layer and a spatial temperature-salinity interpolation field obtained by sparse recovery of the graph signal. Based on the anomaly probability values ​​of each depth layer in the anomaly probability score vectors and the corresponding temperature-salinity gradient amplitudes, the comprehensive anomaly index of the profile is calculated. The comprehensive anomaly index of the profile is substituted into the profile quality adjustment function. The attention temperature parameter of the Transformer encoder in the ocean profile quality assessment and field reconstruction model is adjusted according to the range of the output value of the profile quality adjustment function. Finally, the multi-parameter profile data after quality control and the spatial temperature-salinity interpolation field are written into the data storage and communication module.

6. The marine multi-parameter precision measuring instrument according to claim 5, characterized in that, The TEOS-10 real-time salinity calculation unit based on Horner expansion and Q-format fixed-pointing reorganizes the calculation order of the multivariate polynomials involving corrected temperature values, time-aligned conductivity values, and corrected pressure values ​​in the TEOS-10 international thermodynamic equations using the Horner nested multiplication rule. This ensures that each additional polynomial order requires only one multiplication and one addition, reducing the number of multiplication operations from being proportional to the square of the polynomial order to being linearly proportional to the polynomial order.

7. The marine multi-parameter precision measuring instrument according to claim 6, characterized in that, The Q-format fixed-point formatting scales the corrected temperature value, time-aligned conductivity value, and corrected pressure value to Q16 or Q24 fixed-point format according to their ranges. The intermediate calculation results are executed with 32-bit integer multiplication and addition instructions, and no floating-point arithmetic unit is required throughout the process.

8. The marine multi-parameter precision measuring instrument according to claim 7, characterized in that, The profile comprehensive anomaly index is obtained by normalizing the weighted summation of the anomaly probability values ​​of each depth layer in the anomaly probability scoring vector with the corresponding temperature and salinity gradient magnitude as the weight. The greater the temperature and salinity gradient magnitude, the greater the contribution of the depth layer to the profile comprehensive anomaly index.

9. The marine multi-parameter precision measuring instrument according to claim 8, characterized in that, The profile quality adjustment function is a piecewise linear function with the profile comprehensive anomaly index as the independent variable and the attention temperature adjustment coefficient as the dependent variable. The product of the attention temperature adjustment coefficient and the attention temperature parameter of the Transformer encoder constitutes the updated attention temperature parameter. When the profile comprehensive anomaly index is in the range [0, 0.3), the attention temperature adjustment coefficient takes a first preset value. When the profile comprehensive anomaly index is in the range [0.3, 0.7), the attention temperature adjustment coefficient takes a second preset value. When the profile comprehensive anomaly index is not less than 0.7, the attention temperature adjustment coefficient takes a third preset value. The first preset value is less than the second preset value, and the second preset value is less than the third preset value.

10. The marine multi-parameter precision measuring instrument according to claim 9, characterized in that, The marine profile quality assessment and field reconstruction model consists of two parts: a profile anomaly detection and quality control subnetwork and a multi-platform data fusion and field reconstruction subnetwork. The profile anomaly detection and quality control subnetwork takes the complete temperature, salinity, and depth profile vector, geographic location information, and seasonal coding as inputs. It inputs the complete temperature, salinity, and depth profile vector into a multilayer Transformer encoder with depth as the sequence length. The multilayer Transformer encoder outputs a full profile context representation vector. The full profile context representation vector is followed by an anomaly scoring multilayer sensing head. The anomaly scoring multilayer sensing head outputs an anomaly probability score vector.

Images