Data processing method for a self-contained temperature, salt and depth instrument without disassembly chamber

Abstract

The present invention provides a data processing method for a non-detachable self-contained temperature, salinity, and depth instrument, belonging to the field of electronic digital data processing technology. The method first collects and preprocesses raw data, then uses a variety of mathematical analysis methods to identify instrument errors, correlation errors, drift errors, and environmental interference errors. The intercorrelations between the errors are calculated and an error correlation matrix is ​​constructed. The entropy weight method is used to determine the compensation weights for each error, and targeted corrections are performed using polynomial equations, linear regression equations, Kalman filtering, and temperature gradient compensation equations. Furthermore, a CTSD-NET neural network model is constructed to comprehensively compensate for environmental interference errors. Model parameters are optimized through variable-length window data sampling and cross-validation, and the compensation effect is evaluated to determine the optimal error compensation model. This method achieves high-precision identification and compensation for multiple composite errors in the non-detachable self-contained temperature, salinity, and depth instrument, ensuring the accuracy and reliability of ocean observation data.

Description

Technical Field

[0001] The present invention belongs to the technical field of electronic digital data processing, and in particular relates to a data processing method for a self-contained temperature-salinity-depth instrument without disassembly. Background Art

[0002] In ocean observations, self-contained temperature, salinity, and depth meters are commonly used. They can continuously record key ocean parameters such as water temperature, salinity, and depth over long periods of time without removing the observation platform. Traditional data processing methods primarily use unified calibration coefficients for correction or apply single compensation for specific types of errors based on empirical models, such as correcting temperature drift through curve fitting or eliminating high-frequency noise through frequency domain filtering.

[0003] However, the ocean environment is complex and ever-changing, making existing data processing methods unable to cope with the combined effects of instrument errors, correlation errors, drift errors, and environmental interference errors. Especially under long-term observation conditions, factors such as sensor aging, heat accumulation in the chamber, and ocean turbulence interact, resulting in complex characteristics of nonlinearity, time-varying, and inter-parameter correlation in the measured data. This makes it difficult for traditional single-calibration methods to obtain accurate results.

[0004] Currently, the industry lacks a comprehensive, systematic approach to the coordinated identification and compensation of multiple errors in non-detachable, self-contained temperature, salinity, and depth meters. In particular, there is a lack of data processing technology that can adaptively identify different error types, quantify the interrelationships between errors, and perform targeted compensation. This severely limits the accuracy and reliability of ocean observation data. Specifically, existing technologies present a technical challenge in accurately identifying and compensating for the combined effects of multiple errors during data processing for non-detachable, self-contained temperature, salinity, and depth meters. Summary of the Invention

[0005] In view of this, the present invention provides a data processing method for a non-detachable self-contained temperature, salt and depth meter, which can solve the technical problem in the prior art that the combined influence of multiple errors in the data processing process of the non-detachable self-contained temperature, salt and depth meter is difficult to accurately identify and compensate.

[0006] The present invention is implemented as follows: the present invention provides a data processing method for a non-disassembly-type self-contained temperature and depth meter, comprising: collecting raw data of the salt depth meter and performing sampling and filtering processing; analyzing errors in the raw data, identifying instrument errors, correlation errors, drift errors and environmental interference errors through multiple mathematical analysis methods; calculating the mutual correlation between errors through autocorrelation analysis, constructing an error correlation matrix, and extracting main error influencing factors using a principal component analysis method; determining each error compensation weight based on an entropy weight method, and using corresponding compensation equations for different error types for correction; constructing an error compensation neural network model CTSD-NET, training the neural network model through historical data, and realizing comprehensive compensation for environmental interference errors; randomly changing the data sampling window length, repeatedly executing the error compensation step, and obtaining multiple groups of compensated data under different window lengths as verification sets; quantitatively evaluating the compensation effect, and optimizing the neural network model parameters through a cross-validation method; cyclically executing the verification step multiple times, selecting the optimal result to determine the final error compensation model, and completing the entire data processing process.

[0007] Among them, the raw data of the non-disassembly-type self-contained temperature, salinity and depth instrument include water temperature value, conductivity value, pressure value, attitude angle value, internal battery voltage value, timestamp and internal reference sensor reading; the data sampling window duration is set to a first duration set according to the sensor frequency, and the first duration is 60 to 120 seconds.

[0008] Among them, the analysis of errors in the original data specifically includes: identifying high-frequency noise through Fourier spectrum analysis to determine instrument errors, using the correlation matrix between parameters to calculate and identify correlation errors, using time series trend analysis to detect drift errors, separating environmental interference errors through wavelet transform, and identifying abnormal data points based on the adaptive threshold algorithm.

[0009] Among them, the error compensation weights are determined based on the entropy weight method, and corresponding compensation equations are used for correction according to different error types, specifically including: using a polynomial compensation equation to correct the instrument error, using a linear regression compensation equation to correct the correlation error, using a Kalman filter compensation equation to correct the drift error in real time, and using a temperature gradient compensation equation to correct the measurement error caused by temperature change.

[0010] Among them, constructing an error compensation neural network model CTSD-NET, training the neural network model through historical data, realizing comprehensive compensation for environmental interference errors specifically includes: constructing an error compensation neural network model CTSD-NET, training the error compensation neural network model CTSD-NET through historical data, realizing comprehensive compensation for the environmental interference errors, and using an error compensation equation group to optimize the compensation effect.

[0011] Among them, the error compensation equation group includes an instrument error compensation equation, a correlation error compensation equation, a drift error compensation equation and a temperature gradient compensation equation; the instrument error compensation equation is used to correct the system deviation caused by the inherent accuracy limitation of the equipment; the correlation error compensation equation is used to eliminate the measurement deviation caused by the mutual influence of various parameters; the drift error compensation equation is used to correct the accumulated system deviation during long-term measurement; the temperature gradient compensation equation is used to correct the impact of temperature changes in different water layers on measurement accuracy.

[0012] Among them, the specific structure of the error compensation neural network model CTSD-NET is a deep neural network architecture that combines a multi-layer bidirectional long short-term memory network with a residual connection, and includes an end-to-end learning framework consisting of a data pre-processing layer, a feature extraction layer, an error classification layer, a compensation calculation layer, and a post-processing layer.

[0013] Among them, the feature extraction layer adopts a multi-head attention mechanism to process the complex relationship between different types of errors. The number of attention heads of the multi-head attention mechanism is dynamically adjusted according to the number of error classifications, the number of residual connections is determined according to the length of the data sampling window, and the number of hidden layer neurons is adaptively set according to the sensor accuracy level.

[0014] Among them, the steps of establishing a training dataset in the pre-training process of the error-compensating neural network model CTSD-NET specifically include: screening high-quality data samples with complete labels from historical observation data, constructing a simulated dataset with known error types and amplitudes by artificially adding error factors of varying degrees, and using calibration equipment to collect reference data under a standard environment as true value labels.

[0015] Among them, the training data set is divided into training set, validation set and test set in a ratio of 7:2:1, and the data is standardized to eliminate the dimensional differences between different physical quantities; the entire pre-training process is carried out separately for data under different sea environment conditions to enhance the adaptability of the model and ensure that various error types can be accurately distinguished and compensated in various complex marine environments.

[0016] Compared with the existing technology, the present invention provides a data processing method for a non-disassembly type self-contained temperature, salt and depth meter. The present invention proposes a data processing method for a non-disassembly type self-contained temperature, salt and depth meter based on error classification identification and multi-level compensation. It uses multiple mathematical methods to collaboratively analyze different types of errors and constructs a comprehensive compensation system including the error compensation neural network model CTSD-NET, thereby realizing accurate identification and compensation of various errors in complex marine environments.

[0017] This method accurately distinguishes between inherent instrument errors, inter-parameter correlation errors, drift errors caused by time accumulation, and random environmental interference errors, and adopts targeted compensation strategies based on the error characteristics. Specifically, it constructs an error correlation matrix to quantify the mutual influence of different errors, uses the entropy weight method to determine compensation weights, and employs the CTSD-NET model to achieve intelligent compensation for complex environmental interference, effectively solving the problem of compound errors that is difficult to handle with traditional methods.

[0018] Through multi-time window data verification and cross-validation mechanism, the method of the present invention overcomes the problem of poor adaptability of single error models in the existing technology to complex environments, achieves high precision and stability in data processing, provides more reliable data support for the field of ocean observation, and solves the technical problem that the combined influence of multiple errors is difficult to accurately identify and compensate during the data processing of the self-contained temperature, salinity and depth instrument without disassembly. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, 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 with reference to the accompanying drawings in the embodiments of the present invention.

[0021] like Figure 1 FIG. 1 is a flow chart of a data processing method for a self-contained temperature-salinity-depth instrument provided by the present invention, which includes the following steps:

[0022] S01, collecting raw data from a self-contained temperature, salt and depth meter without disassembly, performing sampling and filtering on the raw data, and setting a data sampling window duration to a first duration according to a sensor frequency;

[0023] S02. Analyze the errors in the raw data of the self-contained temperature, salt and depth instrument without disassembly chamber, identify high-frequency noise by Fourier spectrum analysis to determine instrument errors, calculate and identify correlation errors by using the correlation matrix between parameters, detect drift errors by time series trend analysis, separate environmental interference errors by wavelet transform, and identify abnormal data points based on an adaptive threshold algorithm;

[0024] S03, calculating the mutual correlation between the instrument error, the correlation error, the drift error and the environmental interference error through autocorrelation analysis, constructing an error correlation matrix, and extracting the main error influencing factors using principal component analysis;

[0025] S04. Determine compensation weights for the instrument error, the correlation error, the drift error, and the environmental interference error based on an entropy weight method, correct the instrument error using a polynomial compensation equation, and correct the correlation error using a linear regression compensation equation;

[0026] S05. For the drift error, a Kalman filter compensation equation is used to perform real-time correction, and a temperature gradient compensation equation is used to correct the measurement error caused by temperature change;

[0027] S06, constructing an error compensation neural network model CTSD-NET, training the error compensation neural network model CTSD-NET through historical data to achieve comprehensive compensation for the environmental interference error, and optimizing the compensation effect using an error compensation equation group;

[0028] S07, randomly changing the data sampling window duration to a second duration, repeating steps S02 to S06, and obtaining multiple sets of compensated data with different second durations as validation sets;

[0029] S08. Using root mean square error and relative deviation evaluation indicators to quantitatively evaluate the compensation effect, and optimizing the network parameters of the error compensation neural network model CTSD-NET by cross-validation method;

[0030] S09. Execute steps S07 to S08 repeatedly in a loop, select the optimal results of the root mean square error and the relative deviation evaluation index, determine the final error compensation model and parameter settings, and complete the entire data processing process of the non-disassembly-type self-contained temperature, salt and depth instrument.

[0031] The raw data of the self-contained temperature, salinity and depth meter without disassembly includes water temperature value, conductivity value, pressure value, attitude angle value, internal battery voltage value, timestamp and internal reference sensor reading.

[0032] Among them, the self-contained temperature, salinity and depth meter that does not require disassembly refers to a comprehensive measuring instrument that can measure water temperature, salinity and depth parameters without being removed from the ocean observation equipment and has a built-in data storage unit.

[0033] The data sampling window duration is a first duration that refers to a basic data processing time unit set according to the sensor sampling frequency and the changing characteristics of the ocean environment, and is usually set to 60 to 120 seconds.

[0034] The instrument error refers to the measurement deviation caused by the hardware accuracy limitation of the temperature, salinity and depth instrument itself and the inherent characteristics of electronic components.

[0035] Among them, the correlation error refers to the systematic deviation caused by the mutual influence between different measurement parameters. In the self-contained temperature-salinity-depth instrument with no disassembly chamber, it is mainly caused by the accumulation of heat in the sealed chamber, which leads to the increase of the temperature of the internal electronic components and affects the sensitivity of the sensor.

[0036] The drift error refers to the systematic deviation that gradually accumulates and increases with the extension of measurement time, which is mainly caused by the aging of sensor performance.

[0037] The environmental interference error refers to the random measurement error caused by factors such as ocean environment fluctuations, water turbulence and external electromagnetic interference.

[0038] The second duration refers to a randomly set window length for verifying the stability of the error compensation algorithm, ranging from 0.5 times to 2 times of the first duration.

[0039] Among them, the simple composition of the non-disassembly chamber type self-contained temperature, salt and depth instrument includes a sealed chamber body, a temperature sensor, a conductivity sensor, a pressure sensor, a posture sensor, a data storage module, a power module, an internal reference sensor and a communication interface.

[0040] The root mean square error refers to the square root of the average of the sum of the squares of the differences between the measured value and the true value, and is used to evaluate the accuracy of the measured value.

[0041] The relative deviation evaluation index refers to the ratio of the difference between the measured value and the true value to the true value, expressed as a percentage, and is used to evaluate the accuracy of the measured value.

[0042] The error compensation equation group includes an instrument error compensation equation, a correlation error compensation equation, a drift error compensation equation and a temperature gradient compensation equation;

[0043] The instrument error compensation equation is used to correct the systematic deviation caused by the inherent accuracy limitation of the equipment. The input includes the water temperature value, the conductivity value, the pressure value, the equipment calibration coefficient, the temperature influence factor and the pressure influence factor. The output is the measurement value after instrument error compensation. The measurement value after instrument error compensation is used as the input of the correlation error compensation equation.

[0044] The correlation error compensation equation is used to eliminate the measurement deviation caused by the mutual influence of various parameters. The input includes the measurement value after instrument error compensation, the water temperature value, the conductivity value, the pressure value, the historical correlation coefficient and the environmental factor influence weight. The output is an independent measurement value after eliminating the mutual influence between the parameters. The independent measurement value after eliminating the mutual influence between the parameters is used as the input of the drift error compensation equation;

[0045] The drift error compensation equation is used to correct the accumulated system deviation during the long-term measurement process. The input includes the independent measurement value after eliminating the mutual influence between the parameters, the initial reference value, the timestamp, the equipment aging coefficient and the environmental influence factor. The output is the stable measurement value after eliminating the time accumulation effect. The stable measurement value after eliminating the time accumulation effect is used as the input of the temperature gradient compensation equation;

[0046] The temperature gradient compensation equation is used to correct the impact of temperature changes in different water layers on measurement accuracy. The input includes the stable measurement value after eliminating the time accumulation effect, the vertical temperature distribution, the pressure value, the water body density, the thermodynamic coefficient and the equipment response time. The output is the accurate measurement result after eliminating the influence of the temperature gradient. The accurate measurement result after eliminating the influence of the temperature gradient is used as the input of the error compensation neural network model CTSD-NET.

[0047] The specific structure of the error compensation neural network model CTSD-NET is a deep neural network architecture that combines a multi-layer bidirectional long short-term memory network with a residual connection. It includes an end-to-end learning framework consisting of a data pre-processing layer, a feature extraction layer, an error classification layer, a compensation calculation layer, and a post-processing layer. The feature extraction layer uses a multi-head attention mechanism to process the complex relationship between different types of errors. The number of attention heads of the multi-head attention mechanism is dynamically adjusted according to the number of error classifications, the number of residual connections is determined according to the length of the data sampling window, and the number of hidden layer neurons is adaptively set according to the sensor accuracy level.

[0048] The steps for establishing a training dataset during the pre-training process of the error-compensating neural network model CTSD-NET specifically include first screening high-quality data samples with complete labels from historical observation data, then constructing a simulated dataset with known error types and amplitudes by artificially adding error factors of varying degrees, then using calibration equipment to collect reference data under a standard environment as true value labels, and finally dividing these three parts of data into training, validation, and test sets in a ratio of 7:2:1, and normalizing the data to eliminate dimensional differences between different physical quantities.

[0049] The steps of pre-training the error compensation neural network model CTSD-NET specifically include first initializing the network parameters using the Xavier method to ensure training stability, then using a small batch gradient descent algorithm to train the model while introducing an early stopping mechanism to prevent overfitting, then optimizing the model convergence process through a learning rate decay strategy, then using the cross-validation method to evaluate the model generalization ability and adjust the network hyperparameters based on the verification results, and finally using a test set to evaluate the model performance and save the optimal parameter configuration. The entire pre-training process is carried out separately for data under different marine environmental conditions to enhance the model's adaptability and ensure that various error types can be accurately distinguished and compensated in various complex marine environments.

[0050] The specific implementation of the above steps is described in detail below.

[0051] The specific implementation method of step S01 is to first collect all the original measurement data recorded by the self-contained temperature, salt and depth meter without disassembly, including water temperature value, conductivity value, pressure value, attitude angle value, internal battery voltage value, timestamp and internal reference sensor reading, etc. After the acquisition is completed, the raw data is preliminarily filtered using a Butterworth low-pass filter, and the filter cutoff frequency is set to 0.4 times the sensor sampling frequency to effectively remove high-frequency random noise while retaining the main characteristics of the signal. According to the sensor response characteristics and the ocean environment change cycle, the data sampling window duration is set to the first duration, usually 60 to 120 seconds, which is long enough to capture the short-term change characteristics of water parameters without over-smoothing important details. The purpose of this step is to obtain high-quality initial data to provide a basis for subsequent error analysis.

[0052] The specific implementation of step S02 is to perform multi-dimensional error analysis on the collected raw data. First, the fast Fourier transform algorithm is used to perform spectral analysis on the data, calculate the power spectral density, and identify abnormal peaks with a frequency higher than 0.5 times the sensor's Newquist frequency. These peaks are determined to be noise errors caused by the instrument hardware. Then, a parameter correlation matrix is ​​constructed and the Pearson correlation coefficient between each measurement parameter is calculated. When the absolute value of the correlation coefficient exceeds 0.75, a significant correlation error is determined to exist. Then, a seasonal trend decomposition program is used to perform trend analysis on the time series data. When 20 consecutive data points show monotonic changes and the rate of change exceeds 2 times the standard deviation, it is identified as a drift error. The signal is then decomposed using a multi-scale discrete wavelet transform to extract features of different frequency bands. The portion of the wavelet coefficient that matches the known characteristics of marine environmental changes is identified as environmental interference error. Finally, a threshold is calculated based on the adaptive median absolute deviation method, and data points that deviate from the median by more than 3.5 standard deviations are marked as outliers. The purpose of this step is to comprehensively identify various error sources and lay the foundation for subsequent targeted compensation.

[0053] The specific implementation method of step S03 is to construct an error correlation measurement system. First, the autocorrelation function and cross-correlation function of each error time series are calculated to quantify the degree of mutual influence between different errors. When the absolute value of the cross-correlation coefficient is greater than 0.6, it indicates that there is a significant correlation between the errors. The correlation coefficients of the various errors are then combined to form an error correlation matrix, and the matrix elements are the correlation coefficients between each error pair. The principal component analysis method is then used to perform eigendecomposition on this matrix, and the principal components with eigenvalues ​​greater than 1 are extracted as the main error influencing factors. Usually, the first three principal components can explain more than 85% of the total error variation. The purpose of this step is to quantify the interaction between various errors and provide a basis for subsequent compensation strategies to avoid repeated compensation or insufficient compensation.

[0054] The specific implementation method of step S04 is to establish an error compensation weight system and implement preliminary compensation. First, the compensation weight of each error type is calculated based on the entropy weight method, the information entropy of each type of error is calculated, and the weight coefficient is obtained after normalization. Usually, the instrument error weight is 0.35 to 0.45, the correlation error weight is 0.25 to 0.35, the drift error weight is 0.15 to 0.25, and the environmental interference error weight is 0.1 to 0.2. Then, for the instrument error, a polynomial compensation equation is established to correct the original data. The polynomial order is determined according to the sensor accuracy level, generally 3 to 5 orders. For the correlation error, a multivariate linear regression equation containing interaction terms is established for correction. The regression equation contains each measurement parameter and its cross term, and the coefficient is obtained by fitting the historical calibration data by the least squares method. The purpose of this step is to establish a reasonable error weight distribution mechanism, implement preliminary compensation for systematic errors, and improve the basic quality of the data.

[0055] The specific implementation method of step S05 is to perform real-time correction of dynamic errors. First, for the drift error, a Kalman filter compensation equation is constructed. The state vector contains the measurement value and the drift rate. The process noise covariance matrix and the observation noise covariance matrix are determined by historical data statistics, and the initial state is set by the equipment calibration value. The Kalman filter estimates the drift amount in the current state in real time and makes corrections through two stages: prediction and update. Then, a temperature gradient compensation equation is constructed. The temperature gradient is calculated using the temperature difference between adjacent depth points in the vertical direction. When the absolute value of the temperature gradient exceeds 0.5°C / meter, the measurement deviation caused by the temperature change is calculated and corrected based on the thermal response characteristics of the sensor and the thermodynamic properties of the water body. The purpose of this step is to deal with time-related system errors and measurement deviations caused by changes in ambient temperature, and to further improve data quality.

[0056] The specific implementation of step S06 is to establish a neural network compensation model for complex errors. First, an error compensation neural network model CTSD-NET is constructed. This model adopts a multi-layer bidirectional long short-term memory network structure and combines a residual connection mechanism to enhance gradient propagation. The input layer receives the data processed by the previous steps and the original data. The feature extraction layer uses a multi-head attention mechanism to handle the complex relationship between different error types, with the number of attention heads ranging from 4 to 8. The error classification layer divides the identified errors into different categories and evaluates their intensity. The compensation calculation layer generates the corresponding compensation value. The post-processing layer ensures the physical rationality of the compensation result. The neural network model is then trained using historical calibration data. The loss function adopts a combination of weighted mean square error and smoothed L1 loss. The optimizer uses the Adam algorithm, the learning rate is initialized to 0.001, and a cosine annealing strategy is used for dynamic adjustment. The purpose of this step is to establish an advanced compensation mechanism for environmental interference errors and handle complex nonlinear error patterns that are difficult to capture with linear methods.

[0057] The specific implementation method of step S07 is to verify the stability of the compensation method through data window transformation. First, a plurality of second durations of different lengths are randomly generated, ranging from 0.5 times to 2 times of the first duration, such as 30 to 240 seconds. Then the error analysis and compensation process of steps S02 to S06 are repeated for the data under each second duration window to generate multiple groups of compensated data results. Under different window lengths, the parameter settings need to be adjusted accordingly, such as the short window needs to increase the filtering intensity, and the long window needs to adjust the drift detection threshold. The data processing results for each group of second durations are saved as an independent verification set for subsequent compensation effect evaluation. The purpose of this step is to test the adaptability and stability of the error compensation algorithm under different time scales to ensure that the method is effective under various sampling conditions.

[0058] The specific implementation method of step S08 is to quantitatively evaluate and optimize the compensation effect. First, based on the reference measurement values ​​of the standard calibration equipment, the root mean square error of the data before and after compensation is calculated. The root mean square error calculation formula is the square root of the average of the sum of the squares of the difference between the measured value and the true value. The error reduction rate should reach more than 70%. Then, the relative deviation evaluation index is calculated, that is, the percentage of the difference between the measured value and the true value relative to the true value. The relative deviation should be controlled within 1.5 times the nominal accuracy of the sensor. Then, the k-fold cross-validation method is used, usually with a k value of 5 or 10. The data is divided into k parts, and the k-1 parts are used in turn to train the model. The 1 part is used for verification, and the average performance is finally taken as the evaluation result. Based on the cross-validation results, the network parameters of the error compensation neural network model CTSD-NET, including the number of layers, number of neurons, and attention mechanism parameters, are adjusted to minimize the verification error. The purpose of this step is to objectively evaluate the compensation effect and optimize the model parameters to improve compensation accuracy and generalization ability.

[0059] The specific implementation method of step S09 is to determine the optimal compensation model configuration. First, execute steps S07 to S08 repeatedly, generally not less than 20 times, and use a different random seed to generate the second time window each time to ensure that the test covers multiple time scales. Then perform a statistical analysis on all the loop results, and select the model configuration with the best comprehensive evaluation index of root mean square error and relative deviation. It is usually required that the root mean square error is less than 0.8 times the nominal accuracy of the sensor, and the relative deviation is less than 1%. Finally, the optimal model configuration is solidified into the final error compensation model and parameter settings to complete the full process design of the data processing method of the self-contained temperature, salt and depth instrument without disassembly. The purpose of this step is to screen out the compensation scheme with the best performance through multiple tests to ensure the accuracy and reliability of the data processing results.

[0060] Furthermore, the error compensation neural network model (CTSD-NET) utilizes a deep neural network architecture that combines a multi-layer bidirectional long-short-term memory (LSTM) network with residual connections. The model consists of five functional layers: a data pre-processing layer normalizes and aligns the input data, using the Z-score method to transform all parameters into a standard distribution with a mean of 0 and a standard deviation of 1. The feature extraction layer comprises three to five bidirectional LSTM layers, each containing 64 to 128 neurons. This bidirectional structure simultaneously considers the influence of historical and future data points, capturing temporal relationships. A multi-head attention mechanism is used to handle the complex relationships between different error types. The number of attention heads is typically equal to the number of error classifications plus two, with each attention head focusing on a different feature combination. The error classification layer utilizes a fully connected network structure consisting of two to three hidden layers, each with 0.75 times the number of neurons in the previous layer. A softmax activation function is used to output the probability distribution of each error. The compensation calculation layer generates compensation values ​​based on the probability and magnitude of each error, using a tanh activation function to ensure that the compensation amplitude is within a reasonable range. The post-processing layer combines the original signal with the compensation value through residual connections, applying physical constraints to ensure that the results conform to the actual marine environment. The number of residual connections is determined by the duration of the data sampling window, generally 0.1 times the window length, rounded up. The number of hidden layer neurons is adaptively set based on the sensor accuracy level, with the number of neurons increasing by 25% for each level of accuracy increase.

[0061] Furthermore, the training dataset construction step first screens high-quality data samples from the historical observation database that have undergone at least three standard calibrations and have a data integrity of over 98%. The number of samples is no less than 1,000, covering different sea areas and seasonal conditions. Then, a simulated dataset with known error types and amplitudes is constructed by artificially adding error factors of varying degrees. This includes adding Gaussian white noise to simulate instrument error, with coefficients ranging from 0.5 to 2 times the standard deviation; adding linear and exponential drift to simulate drift error, with a drift rate of 0.01% to 0.1% / hour; adding cross-influence terms to simulate correlation errors; and adding periodic fluctuations to simulate environmental interference. Third, reference data is collected using calibration equipment in a standard environment as true value labels. The calibration environment temperature is controlled at 20±0.1°C, the concentration error of the salinity standard solution is less than 0.001%, and the pressure standard accuracy reaches 0.01% of full scale. Finally, the three parts of data are divided into training set, validation set and test set in a ratio of 7:2:1, and the data are standardized to eliminate the dimensional differences between different physical quantities. The standardization uses the minimum-maximum scaling method to linearly map each parameter to the interval [0, 1].

[0062] The mathematical model or calculation process involved in the present invention is described in detail below.

[0063] In step S02, the original data is subjected to spectrum analysis using fast Fourier transform to calculate the power spectrum density, which is specifically expressed as follows:

[0064] ;

[0065] Where, is the power spectral density; is the original data time series; is the number of data points; is the frequency; Is an imaginary unit.

[0066] The power spectral density calculation process specifically includes: first, segmenting the time series data, with each segment length being the first time length of the data sampling window, and adjacent segments overlapping by 50%; then applying the Hanning window function to each segment of data to reduce spectral leakage; then calculating the fast Fourier transform of each segment of data; finally, averaging the transformed results of all segments to obtain a smoothed power spectral density estimate.

[0067] In step S02, a parameter correlation matrix is ​​constructed and the Pearson correlation coefficient is calculated, which is specifically expressed as follows:

[0068] ;

[0069] Where, is the parameter correlation matrix; For the Parameters and Pearson correlation coefficient between the parameters; is the total number of parameters.

[0070] The calculation formula of Pearson correlation coefficient is:

[0071] ;

[0072] Where, For the The first parameter observations; For the The average value of the parameters; For the The first parameter observations; For the The average value of the parameters; is the number of observation data points.

[0073] In step S02, the signal is decomposed using multi-scale discrete wavelet transform, which is specifically expressed as follows:

[0074] ;

[0075] Where, is the wavelet coefficient; is the original signal; is the wavelet mother function; is the scale parameter; is the translation parameter.

[0076] The specific implementation of wavelet transform is through the Mallat algorithm, which includes: first selecting a suitable wavelet mother function, usually the Daubechies wavelet; then determining the number of decomposition layers, generally 3 to 5 layers; then obtaining detail coefficients and approximate coefficients through high-pass filters and low-pass filters respectively; finally, reconstructing the signal components based on the coefficients of each layer.

[0077] In step S02, the threshold is calculated using the adaptive median absolute deviation method, which is specifically expressed as follows:

[0078] ;

[0079] Where, is the outlier detection threshold; is the median of the data; is the median absolute deviation; is the amplification factor, which is usually taken as 3.5.

[0080] The formula for calculating the median absolute deviation is:

[0081] ;

[0082] Where, is a data point; For the entire dataset; Calculates the median function.

[0083] In step S03, the autocorrelation function and the cross-correlation function are calculated, which are specifically expressed as follows:

[0084] ;

[0085] ;

[0086] Where, For time series The autocorrelation function of For time series and The cross-correlation function of For time delay; is the data length; and Sequence and The mean of .

[0087] In step S03, the error correlation matrix is ​​constructed as follows:

[0088] ;

[0089] Where, is the error correlation matrix; For the Class error and The value of the cross-correlation function of the class error at zero delay; the subscripts 1, 2, 3, and 4 represent instrument error, correlation error, drift error, and environmental interference error, respectively.

[0090] In step S03, the principal component analysis method is used to perform eigendecomposition on the error correlation matrix, which is specifically expressed as follows:

[0091] ;

[0092] Where, is the error correlation matrix; is the eigenvector matrix; is a diagonal matrix of eigenvalues.

[0093] The calculation process of the principal components is as follows: first calculate the eigenvalues ​​and eigenvectors of the error correlation matrix; then sort them from large to small according to the eigenvalues; then determine the number of principal components based on the cumulative contribution rate, usually selecting the first few principal components with a cumulative contribution rate of more than 85%; finally, use the eigenvectors to construct the principal components.

[0094] In step S04, the entropy weight method is used to calculate the compensation weight of each error type, which is specifically expressed as follows:

[0095] ;

[0096] ;

[0097] Where, For the Information entropy of class errors; For the The first observation data point The normalized value of the class error; is the number of observation data points; For the The weight of the class error; is the number of error types. .

[0098] The calculation formula is:

[0099] ;

[0100] Where, For the The first observation data point The raw value of the class error.

[0101] In step S04, the polynomial compensation equation is used to correct the instrument error, which is specifically expressed as follows:

[0102] ;

[0103] Where, is the corrected measurement value; is the original measurement value; is the water temperature; is the pressure value; are the polynomial coefficients; is the temperature influence coefficient; is the pressure influence coefficient; and are the interaction coefficients of temperature and pressure, respectively; is the polynomial order, usually 3 to 5; and is the order of the interaction term, which is generally 2 to 3.

[0104] The above parameters are obtained by least squares fitting and need to be trained using standard calibration data. The calibration data should cover the temperature, pressure and measurement value combinations within the operating range of the device.

[0105] In step S04, the multivariate linear regression equation corrects the correlation error, which is specifically expressed as follows:

[0106] ;

[0107] Where, is the corrected parameter value; is the parameter value after instrument error compensation; 、 、 are the rates of change of temperature, conductivity, and pressure, respectively; 、 、 They are the current temperature, conductivity and pressure values ​​respectively; 、 、 is the influence coefficient of the rate of change; 、 、 is the parameter interaction coefficient; is the three-parameter interaction coefficient.

[0108] Each coefficient is obtained by fitting historical calibration data using the multivariate linear regression method, which requires a large number of experimental measurements of the measurement deviations under different parameter combinations.

[0109] In step S05, the Kalman filter compensation equation corrects the drift error in real time. The state space model is specifically expressed as follows:

[0110] ;

[0111] ;

[0112] Where, is the state vector, which contains the measurement value and drift rate; is the state transfer matrix; is process noise, with mean zero and covariance of Gaussian distribution; is the observation vector; is the observation matrix; is the observation noise, which has a mean of zero and a covariance of Gaussian distribution.

[0113] Specifically, the state vector and matrix are defined as:

[0114] ; ; ;

[0115] Where, is the measured value at the current moment; is the drift rate; is the sampling time interval.

[0116] The prediction and update process of Kalman filter is:

[0117] predict:

[0118] ;

[0119] ;

[0120] renew:

[0121] ;

[0122] ;

[0123] ;

[0124] Where, is the prior state estimate; is the prior estimate error covariance; is the Kalman gain; is the posterior state estimate; is the posterior estimated error covariance; is the identity matrix.

[0125] Process noise covariance matrix and the observation noise covariance matrix Determined through historical data statistics, Reflects the uncertainty of the system state changing over time, Reflects the uncertainty of the measurement process. Initial state The initial covariance matrix is ​​set by the device calibration value. Typically set to a large value to reflect the uncertainty of the initial estimate.

[0126] In step S05, the temperature gradient compensation equation is specifically expressed as follows:

[0127] ;

[0128] Where, is the corrected temperature value; To measure the temperature value; is the thermal response coefficient, ranging from 0.05 to 0.15; is the vertical temperature gradient; is the sensor response time; is the environmental response time.

[0129] The calculation formula for the vertical temperature gradient is:

[0130] ;

[0131] Where, and are the temperature values ​​of the upper and lower adjacent depth points respectively; is the depth interval.

[0132] Sensor response time Provided by the manufacturer or obtained through experimental measurement, the environmental response time Determined by the thermodynamic properties of water, it is usually 0.2 to 0.5 times the sensor response time.

[0133] In step S08, the calculation formula of the root mean square error is:

[0134] ;

[0135] Where, is the root mean square error; is the measured value; is the true value; is the number of data points.

[0136] The calculation formula of the relative deviation evaluation index is:

[0137] ;

[0138] Where, is the relative deviation percentage; is the measured value; is the true value; is the number of data points.

[0139] Optionally, in step S02, a seasonal trend decomposition program performs trend analysis on the time series data, and its mathematical expression is as follows:

[0140] ;

[0141] Where, is the original time series; is the trend component; is the seasonal component; is the residual component.

[0142] The trend component is extracted by the moving average method:

[0143] ;

[0144] Where, is the half-length of the moving average window, which is usually a multiple of the observation frequency.

[0145] The criteria for determining drift error are:

[0146] ;

[0147] Where, is the standard deviation of the residual component; is the number of continuous data points. ; is a symbolic function.

[0148] Optionally, in step S06, the error compensation neural network model loss function adopts a combination of weighted mean square error and smooth L1 loss, which is specifically expressed as follows:

[0149] ;

[0150] Where, is the total loss function; is the weight coefficient, ranging from 0.6 to 0.8; is the mean square error; is the smooth L1 loss.

[0151] The formula for calculating the mean square error is:

[0152] ;

[0153] The smooth L1 loss calculation formula is:

[0154] ;

[0155] Where, ; is the true value; is the model prediction value; is the sample size.

[0156] Specifically, the present invention is based on the concepts of error separation and identification and multi-level compensation. Through systematic analysis of various error characteristics, a complete processing framework for error identification, correlation analysis, and hierarchical compensation is constructed. First, multi-dimensional analysis methods such as Fourier spectrum analysis, correlation matrix calculation, time series trend analysis, and wavelet transform are used to accurately identify and differentiate instrument errors, correlation errors, drift errors, and environmental interference errors, laying the foundation for subsequent targeted compensation.

[0157] Secondly, the present invention quantifies the intercorrelations between different errors through autocorrelation analysis and principal component analysis, constructs an error correlation matrix, and reveals the complex interaction mechanisms between errors. Based on this understanding, the entropy weight method is used to determine the compensation weights for various types of errors, making the compensation process more consistent with the actual error contribution ratio and avoiding the problems of under-compensation or over-compensation in traditional methods. For different error types, targeted corrections are made using mechanisms such as polynomial compensation equations, linear regression compensation equations, Kalman filtering, and temperature gradient compensation equations.

[0158] The core innovation of this invention lies in the construction of a CTSD-NET neural network model that integrates a multi-head attention mechanism and residual connections. This model can learn the nonlinear representation and time-varying characteristics of complex environmental interference. The temporal dependencies of errors are captured by a bidirectional long short-term memory network, and the original signal characteristics are preserved using residual connections to achieve intelligent compensation for environmental interference errors. The use of a variable-length window verification strategy and a cross-validation optimization mechanism further enhances the model's generalization and robustness, ensuring the stability and consistency of the compensation effect under different marine environmental conditions.

[0159] Overall, the present invention achieves high-precision compensation for complex errors of the self-contained temperature-salinity-depth meter without disassembly of the chamber through a systematic approach of precise error identification, quantitative analysis of correlation, hierarchical compensation strategy and integration of intelligent models, solving the technical problem that traditional single compensation methods are difficult to cope with the compound influence of errors.

[0160] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0161] The specific implementation method of step S01 is to first collect all the original measurement data recorded by the self-contained temperature, salt and depth meter without disassembly, including water temperature value, conductivity value, pressure value, attitude angle value, internal battery voltage value, timestamp and internal reference sensor reading, etc. After the acquisition is completed, the raw data is preliminarily filtered using a Butterworth low-pass filter, and the filter cutoff frequency is set to 0.4 times the sensor sampling frequency to effectively remove high-frequency random noise while retaining the main characteristics of the signal. According to the sensor response characteristics and the ocean environment change cycle, the data sampling window duration is set to the first duration, usually 60 to 120 seconds, which is long enough to capture the short-term change characteristics of water parameters without over-smoothing important details. The purpose of this step is to obtain high-quality initial data to provide a basis for subsequent error analysis.

[0162] The specific implementation of step S02 is to perform multi-dimensional error analysis on the collected raw data. First, the fast Fourier transform algorithm is used to perform spectrum analysis on the data, calculate the power spectrum density, and identify abnormal peaks with a frequency higher than 0.5 times the sensor's Newquist frequency. These peaks are determined to be noise errors caused by the instrument hardware. The power spectrum density calculation formula is:

[0163] ;

[0164] Where, is the power spectral density; is the original data time series; is the number of data points; is the frequency; Is an imaginary unit.

[0165] The power spectral density calculation process specifically includes: first, segmenting the time series data, with each segment length being the first time length of the data sampling window, and adjacent segments overlapping by 50%; then applying the Hanning window function to each segment of data to reduce spectral leakage; then calculating the fast Fourier transform of each segment of data; finally, averaging the transformed results of all segments to obtain a smoothed power spectral density estimate.

[0166] Then, the parameter correlation matrix is ​​constructed and the Pearson correlation coefficient between each measurement parameter is calculated. When the absolute value of the correlation coefficient exceeds 0.75, it is determined that there is a significant correlation error. The parameter correlation matrix is ​​expressed as:

[0167] ;

[0168] Where, is the parameter correlation matrix; For the Parameters and Pearson correlation coefficient between the parameters; is the total number of parameters.

[0169] The formula for calculating the Pearson correlation coefficient is:

[0170] ;

[0171] Where, For the The first parameter observations; For the The average value of the parameters; For the The first parameter observations; For the The average value of the parameters; is the number of observation data points.

[0172] Then, the seasonal trend decomposition procedure is used to perform trend analysis on the time series data. When 20 consecutive data points show monotonic changes and the rate of change exceeds 2 times the standard deviation, it is identified as a drift error. The mathematical expression of the seasonal trend decomposition procedure is:

[0173] ;

[0174] Where, is the original time series; is the trend component; is the seasonal component; is the residual component.

[0175] The trend component is extracted by the moving average method:

[0176] ;

[0177] Where, is the half-length of the moving average window, which is usually a multiple of the observation frequency.

[0178] The criteria for determining drift error are:

[0179] ;

[0180] Where, is the standard deviation of the residual component; is the number of continuous data points. ; is a symbolic function.

[0181] Then, the signal is decomposed by multi-scale discrete wavelet transform, the characteristics of different frequency bands are extracted, and the part of the wavelet coefficient that matches the known ocean environment change characteristics is identified as environmental interference error. The calculation formula of wavelet transform is:

[0182] ;

[0183] Where, is the wavelet coefficient; is the original signal; is the wavelet mother function; is the scale parameter; is the translation parameter.

[0184] Finally, the threshold is calculated based on the adaptive median absolute deviation method, and data points that deviate from the median by more than 3.5 standard deviations are marked as outliers. The adaptive threshold calculation formula is:

[0185] ;

[0186] Where, is the outlier detection threshold; is the median of the data; is the median absolute deviation; is the amplification factor, which is usually taken as 3.5.

[0187] The formula for calculating the median absolute deviation is:

[0188] ;

[0189] Where, is a data point; For the entire dataset; Calculates the median function.

[0190] The purpose of this step is to comprehensively identify various error sources and lay the foundation for subsequent targeted compensation.

[0191] The specific implementation of step S03 is to construct an error correlation measurement system. First, the autocorrelation function and cross-correlation function of each error time series are calculated to quantify the degree of mutual influence between different errors. The calculation formula of the autocorrelation function and cross-correlation function is:

[0192] ;

[0193] ;

[0194] Where, For time series The autocorrelation function of For time series and The cross-correlation function of For time delay; is the data length; and Sequence and The mean of .

[0195] When the absolute value of the correlation coefficient is greater than 0.6, it indicates that there is a significant correlation between the errors. The correlation coefficients of various types of errors are then combined into an error correlation matrix, where the matrix elements are the correlation coefficients between each pair of errors. The error correlation matrix is ​​expressed as:

[0196] ;

[0197] Where, is the error correlation matrix; For the Class error and The value of the cross-correlation function of the class error at zero delay; the subscripts 1, 2, 3, and 4 represent instrument error, correlation error, drift error, and environmental interference error, respectively.

[0198] Then, the principal component analysis method is used to perform eigendecomposition on this matrix, and the principal components with eigenvalues ​​greater than 1 are extracted as the main error influencing factors. Usually, the first three principal components can explain more than 85% of the total error variation. The calculation formula of principal component analysis is:

[0199] ;

[0200] Where, is the error correlation matrix; is the eigenvector matrix; is a diagonal matrix of eigenvalues.

[0201] The purpose of this step is to quantify the interactions between various types of errors and provide a basis for subsequent compensation strategies to avoid repeated compensation or insufficient compensation.

[0202] The specific implementation of step S04 is to establish an error compensation weight system and implement preliminary compensation. First, the compensation weight of each error type is calculated based on the entropy weight method, the information entropy of each error is calculated, and the weight coefficient is obtained after normalization. The calculation formula of the entropy weight method is:

[0203] ;

[0204] ;

[0205] Where, For the Information entropy of class errors; For the The first observation data point The normalized value of the class error; is the number of observation data points; For the The weight of the class error; is the number of error types. .

[0206] The calculation formula is:

[0207] ;

[0208] Where, For the The first observation data point The raw value of the class error.

[0209] Typically, the instrument error weight is 0.35 to 0.45, the correlation error weight is 0.25 to 0.35, the drift error weight is 0.15 to 0.25, and the environmental interference error weight is 0.1 to 0.2. A polynomial compensation equation is then established to correct the raw data for the instrument error. The polynomial order is determined by the sensor accuracy level and is generally 3 to 5. The polynomial compensation equation is in the form of:

[0210] ;

[0211] Where, is the corrected measurement value; is the original measurement value; is the water temperature; is the pressure value; are the polynomial coefficients; is the temperature influence coefficient; is the pressure influence coefficient; and are the interaction coefficients of temperature and pressure, respectively; is the polynomial order, usually 3 to 5; and is the order of the interaction term, which is generally 2 to 3.

[0212] For the correlation error, a multiple linear regression equation containing interaction terms is established for correction. The regression equation contains each measurement parameter and its cross term, and the coefficients are obtained by fitting the historical calibration data using the least squares method. The form of the multiple linear regression equation is:

[0213] ;

[0214] Where, is the corrected parameter value; is the parameter value after instrument error compensation; 、 、 are the rates of change of temperature, conductivity, and pressure, respectively; 、 、 They are the current temperature, conductivity and pressure values ​​respectively; 、 、 is the influence coefficient of the rate of change; 、 、 is the parameter interaction coefficient; is the three-parameter interaction coefficient.

[0215] The purpose of this step is to establish a reasonable error weight distribution mechanism, implement preliminary compensation for systematic errors, and improve the basic quality of the data.

[0216] The specific implementation of step S05 is to perform real-time correction of dynamic errors. First, for the drift error, a Kalman filter compensation equation is constructed. The state vector contains the measurement value and the drift rate. The process noise covariance matrix and the observation noise covariance matrix are determined by historical data statistics. The initial state is set by the equipment calibration value. The Kalman filter state space model is:

[0217] ;

[0218] ;

[0219] Where, is the state vector, which contains the measurement value and drift rate; is the state transfer matrix; is process noise, with mean zero and covariance of Gaussian distribution; is the observation vector; is the observation matrix; is the observation noise, which has a mean of zero and a covariance of Gaussian distribution.

[0220] Specifically, the state vector and matrix are defined as:

[0221] ;

[0222] ;

[0223] ;

[0224] Where, is the measured value at the current moment; is the drift rate; is the sampling time interval.

[0225] The Kalman filter estimates the drift in the current state and makes corrections in real time through two stages: prediction and update:

[0226] predict:

[0227] ;

[0228] ;

[0229] renew:

[0230] ;

[0231] ;

[0232] ;

[0233] Where, is the prior state estimate; is the prior estimate error covariance; is the Kalman gain; is the posterior state estimate; is the posterior estimated error covariance; is the identity matrix.

[0234] Then, a temperature gradient compensation equation is constructed, and the temperature gradient is calculated using the temperature difference between adjacent depth points in the vertical direction. When the absolute value of the temperature gradient exceeds 0.5°C / meter, the measurement deviation caused by the temperature change is calculated and corrected based on the thermal response characteristics of the sensor and the thermodynamic properties of the water. The temperature gradient compensation equation is in the form of:

[0235] ;

[0236] Where, is the corrected temperature value; To measure the temperature value; is the thermal response coefficient, ranging from 0.05 to 0.15; is the vertical temperature gradient; is the sensor response time; is the environmental response time.

[0237] The calculation formula for the vertical temperature gradient is:

[0238] ;

[0239] Where, and are the temperature values ​​of the upper and lower adjacent depth points respectively; is the depth interval.

[0240] The purpose of this step is to deal with time-related systematic errors and measurement deviations caused by ambient temperature changes, further improving data quality.

[0241] The specific implementation method of step S06 is to establish a neural network compensation model for complex errors. First, an error compensation neural network model CTSD-NET is constructed. The model adopts a multi-layer bidirectional long short-term memory network structure and combines a residual connection mechanism to enhance gradient propagation. The input layer receives the data processed by the previous steps and the original data. The feature extraction layer uses a multi-head attention mechanism to process the complex relationship between different error types, and the number of attention heads is 4 to 8. The error classification layer divides the identified errors into different categories and evaluates their intensity. The compensation calculation layer generates corresponding compensation values, and the post-processing layer ensures the physical rationality of the compensation results. The neural network model is then trained using historical calibration data, and the loss function adopts a combination of weighted mean square error and smoothed L1 loss:

[0242] ;

[0243] Where, is the total loss function; is the weight coefficient, ranging from 0.6 to 0.8; is the mean square error; is the smooth L1 loss.

[0244] The formula for calculating the mean square error is:

[0245] ;

[0246] The smooth L1 loss calculation formula is:

[0247] ;

[0248] Where, ; is the true value; is the model prediction value; is the sample size.

[0249] The optimizer uses the Adam algorithm, with an initial learning rate of 0.001 and a cosine annealing strategy for dynamic adjustment. The purpose of this step is to establish an advanced compensation mechanism for environmental interference errors and handle complex nonlinear error patterns that are difficult to capture with linear methods.

[0250] The specific implementation method of step S07 is to verify the stability of the compensation method through data window transformation. First, a plurality of second durations of different lengths are randomly generated, ranging from 0.5 times to 2 times of the first duration, such as 30 to 240 seconds. Then the error analysis and compensation process of steps S02 to S06 are repeated for the data under each second duration window to generate multiple groups of compensated data results. Under different window lengths, the parameter settings need to be adjusted accordingly, such as the short window needs to increase the filtering intensity, and the long window needs to adjust the drift detection threshold. The data processing results for each group of second durations are saved as an independent verification set for subsequent compensation effect evaluation. The purpose of this step is to test the adaptability and stability of the error compensation algorithm under different time scales to ensure that the method is effective under various sampling conditions.

[0251] The specific implementation of step S08 is to quantitatively evaluate and optimize the compensation effect. First, based on the reference measurement value of the standard calibration equipment, the root mean square error of the data before and after compensation is calculated:

[0252] ;

[0253] Where, is the root mean square error; is the measured value; is the true value; is the number of data points.

[0254] The error reduction rate should reach more than 70%. Then calculate the relative deviation evaluation index:

[0255] ;

[0256] Where, is the relative deviation percentage; is the measured value; is the true value; is the number of data points.

[0257] The relative deviation should be within 1.5 times the nominal accuracy of the sensor. Next, a k-fold cross-validation method is used, typically with a k value of 5 or 10. The data is divided into k parts. The k-1 parts are used in turn to train the model, and the 1 part is used for validation. The average performance is ultimately used as the evaluation result. Based on the cross-validation results, the network parameters of the error compensation neural network model CTSD-NET, including the number of layers, number of neurons, and attention mechanism parameters, are adjusted to minimize the validation error. The purpose of this step is to objectively evaluate the compensation effect and optimize the model parameters to improve compensation accuracy and generalization ability.

[0258] The specific implementation method of step S09 is to determine the optimal compensation model configuration. First, execute steps S07 to S08 repeatedly, generally not less than 20 times, and use a different random seed to generate the second time window each time to ensure that the test covers multiple time scales. Then perform a statistical analysis on all the loop results, and select the model configuration with the best comprehensive evaluation index of root mean square error and relative deviation. It is usually required that the root mean square error is less than 0.8 times the nominal accuracy of the sensor, and the relative deviation is less than 1%. Finally, the optimal model configuration is solidified into the final error compensation model and parameter settings to complete the full process design of the data processing method of the self-contained temperature, salt and depth instrument without disassembly. The purpose of this step is to screen out the compensation scheme with the best performance through multiple tests to ensure the accuracy and reliability of the data processing results.

[0259] The core advantage of the non-disassembly-type, self-contained temperature, salinity, and depth instrument described in Example 1 is that it can collect and store data without removing the instrument from the observation platform. The instrument features an integrated design, consisting of a sealed chamber, a sensor system, a data processing unit, a storage module, and a power supply system. The sealed chamber is constructed from a high-strength titanium alloy, offering excellent pressure and corrosion resistance, capable of withstanding environmental pressures at depths of 6,000 meters. The sensor system includes a high-precision platinum resistance temperature sensor, a four-electrode conductivity sensor, a piezoresistive pressure sensor, and a nine-axis attitude sensor. The data processing unit utilizes a low-power microprocessor, integrating data acquisition, filtering, and error compensation algorithms. The storage module utilizes a high-capacity flash memory chip, enabling continuous operation and recording of up to one year of observation data. The power supply system utilizes a high-energy-density lithium battery pack equipped with intelligent power management circuitry to extend operating time. The structural design of the non-disassembly-type, self-contained temperature, salinity, and depth instrument utilizes precise connections between various components to form a fully functional system. The high-strength titanium alloy sealed chamber is internally divided into compartments. The main compartment houses the electronics, while the sensor compartment is located at the front of the instrument. Each sensor is connected to the data processing unit via a waterproof connector. The temperature and conductivity sensors are mounted on external probes, the pressure sensor is embedded in the chamber wall, and the attitude sensor is fixed to the main circuit board. The data processing unit is connected to the storage module via a high-speed bus, enabling real-time data writing. The power supply system provides power to all components via a voltage-stabilizing circuit, while the power management chip on the main circuit board controls the power supply status of each module. A watertight communication interface is installed on the outside of the instrument, allowing it to be connected to a computer for data reading without opening the chamber. A pressure-balancing valve is also provided to accommodate depth fluctuations. All electronic components are secured with shock-absorbing brackets to ensure stable operation in complex marine environments.

[0260] The internal reference sensor is a key component of the chamber-free, self-contained temperature, salinity, and depth instrument, providing system internal operating status monitoring and parameter calibration. It consists of a temperature reference sensor, a voltage reference sensor, and a clock reference sensor. The temperature reference sensor utilizes a high-stability semiconductor temperature sensor to monitor internal instrument temperature changes in real time, providing baseline data for the temperature compensation algorithm. The voltage reference sensor utilizes a high-precision bandgap reference source to monitor system power supply voltage fluctuations, ensuring that the measurement circuit operates in a stable voltage environment. The clock reference sensor utilizes a temperature-compensated crystal oscillator to provide a precise time reference, ensuring the accuracy of data sampling. The internal reference sensor's output signal undergoes analog-to-digital conversion and is stored alongside the environmental parameter measurement data for use in error compensation calculations during subsequent data processing.

[0261] The workflow of the non-disassembly-type, self-contained temperature, salinity, and depth meter begins with a pressure sensor detecting depth changes, triggering a measurement. Temperature and conductivity sensors then simultaneously collect water parameters. The raw data undergoes analog-to-digital conversion and enters the data processing unit. A microprocessor then performs a preliminary filtering algorithm to eliminate high-frequency noise and performs preliminary error compensation based on internal reference sensor data. The processed data is stored in flash memory at preset intervals, while the raw data is also preserved for more accurate error analysis and compensation. During long-term operation, the instrument's internal adaptive power management system dynamically adjusts the circuit operating mode based on the measurement frequency to maximize battery life.

[0262] To adapt to different marine environmental conditions, the non-disassembly-type self-contained temperature, salinity and depth instrument provides multiple sampling mode settings. In standard mode, temperature measurement accuracy is better than ±0.002°C, conductivity measurement accuracy is better than ±0.003mS / cm, and pressure measurement accuracy is better than 0.05% of the full scale. In high-precision mode, measurement accuracy is further improved by increasing the number of sampling points and enabling more complex digital filtering algorithms. Fast response mode is suitable for vertical profile observations and can achieve a high-frequency sampling rate of 24 times per second. In low-power mode, the instrument operates at a lower frequency, extending the standby time to more than three times that of normal mode. All mode parameters can be flexibly configured through the communication interface to meet the requirements of different observation tasks.

[0263] To better understand and implement the present invention, Example 2 of a specific application scenario of the present invention is provided below: During a certain oceanographic survey voyage, researchers used a self-contained temperature, salinity, and depth instrument, which required no disassembly, to conduct 30 days of continuous observations at a depth of approximately 2,000 meters, covering various marine environmental conditions. The instrument was equipped with a temperature sensor (accuracy of ±0.002°C), a conductivity sensor (accuracy of ±0.003mS / cm), a pressure sensor (accuracy of ±0.1% of full scale), and an attitude sensor, with a sampling frequency of 1Hz. During the instrument's long-term operation, the raw data exhibited a certain degree of error due to factors such as instrument errors, environmental interference, and sensor performance degradation. The researchers processed the raw data using the data processing method of the present invention to improve data quality.

[0264] First, based on the sensor response characteristics and the periodicity of ocean environmental changes, the researchers set the data sampling window duration to 90 seconds. The raw data was then processed using a Butterworth low-pass filter with a filter cutoff frequency set to 0.4 times the sensor sampling frequency, or 0.4 Hz. This filter parameter setting effectively removes high-frequency random noise from the self-contained temperature, salinity, and depth meter acquisition process while preserving the key characteristics of water parameter changes. The statistical characteristics of the raw and filtered data are shown in Table 1:

[0265] Table 1 Comparison of statistical characteristics of original data and filtered data

[0266]

[0267] Next, the researchers conducted a multi-dimensional error analysis on the filtered data. By analyzing the power spectrum of the temperature data using the fast Fourier transform algorithm, it was found that there was an obvious abnormal peak in the 0.32Hz frequency band, which was confirmed to be a noise error caused by the temperature sensor hardware. By constructing a parameter correlation matrix, the Pearson correlation coefficient between temperature and conductivity was calculated to be 0.83, and the correlation coefficient between temperature and pressure was 0.61, indicating that there was a significant correlation error. The seasonal trend decomposition program was applied to analyze the conductivity data, and it was found that it showed a monotonic upward trend during the 20-hour observation period, with a rate of change of 2.5 times the standard deviation, and it was determined that there was a drift error. The pressure signal was decomposed by multi-scale discrete wavelet transform, and the environmental interference error that matched the tidal cycle (about 12 hours) was identified. Finally, the outlier detection threshold was set by the adaptive median absolute deviation method, and 32 abnormal data points were marked. The statistics of various error characteristics are shown in Table 2:

[0268] Table 2 Statistics of various error characteristics

[0269]

[0270] When constructing the error correlation measurement system, the researchers calculated the autocorrelation and cross-correlation functions of various error time series to obtain an error correlation matrix. Principal component analysis was then used to extract the main error influencing factors. The first three principal components explained 88.4% of the total error variation. The first principal component was primarily composed of instrument error and correlation error, the second principal component was primarily composed of drift error, and the third principal component was primarily composed of environmental interference error. Based on the principal component analysis results and the entropy weight method, the compensation weights for each error type were calculated as follows: 0.39 for instrument error, 0.32 for correlation error, 0.19 for drift error, and 0.10 for environmental interference error.

[0271] Based on the calculated weights, the researchers first compensated for the instrument error. A fourth-order polynomial compensation equation was obtained by fitting the standard calibration data to correct the temperature data:

[0272] ;

[0273] in, is the corrected temperature value, is the original temperature value, is the pressure value.

[0274] For the associated errors, the researchers established a multiple linear regression equation including interaction terms:

[0275] ;

[0276] in, is the corrected conductivity value, is the conductivity value after instrument error compensation, 、 、 are the rates of change of temperature, conductivity and pressure, respectively, 、 、 They are the current temperature, conductivity and pressure values ​​respectively.

[0277] For drift error, the researchers constructed a Kalman filter compensation equation, where the state vector contains the conductivity measurement value and the drift rate. and the observation noise covariance matrix Determined through historical data statistics:

[0278] ;

[0279] ;

[0280] Initial state vector and the initial covariance matrix Set to:

[0281] ;

[0282] ;

[0283] The Kalman filter dynamically estimates and corrects the conductivity drift error during real-time processing. In areas with large temperature gradients (temperature gradients in the thermocline reach 1.2°C / m), the temperature gradient compensation equation is applied to correct the measured data. The compensation parameter is set to , Second, Second.

[0284] For the environmental interference errors that are difficult to handle with classical methods, researchers have constructed an error compensation neural network model CTSD-NET, which consists of a three-layer bidirectional long short-term memory network with 96 neurons in each layer, combined with six residual connections and a six-head attention mechanism. The model is trained using historical calibration data, and the loss function weight coefficient The learning rate is set to 0.7, the initial value is 0.001, and the cosine annealing strategy is used to dynamically adjust during the training process. The training dataset, validation set, and test set configuration are shown in Table 3:

[0285] Table 3 CTSD-NET model training data configuration table

[0286]

[0287] To verify the stability of the compensation method, the researchers randomly changed the data sampling window length to 45 seconds, 70 seconds, 110 seconds, 150 seconds and 180 seconds, and repeated the error analysis and compensation process for the data of each window length to generate multiple sets of compensation results as verification sets. The 5-fold cross-validation method was used to evaluate the compensation effect, and the CTSD-NET network parameters were optimized based on the verification results, especially the number of attention mechanism heads and the number of residual connections. After 25 cycles of testing, the model configuration with the best comprehensive root mean square error and relative deviation was selected as the final error compensation model. The comparison of data quality indicators before and after compensation is shown in Table 4:

[0288] Table 4 Comparison of quality indicators before and after data processing

[0289]

[0290] Traditional ocean observation data processing methods primarily use single mathematical models or statistical methods for error compensation, such as polynomial fitting, moving average filtering, and Kalman filtering. These methods are unable to fully identify and address the interactive effects of multiple error factors in complex ocean environments. These methods typically treat errors as independent entities, ignoring the interrelationships between different errors. This results in limited compensation effectiveness, especially in long-term observations and complex ocean environments. For example, traditional methods struggle to simultaneously address the compound errors caused by instrument drift over time and sudden changes in ambient temperature. The relative deviation of compensated data is typically between 0.02% and 0.05%, which cannot meet the needs of high-precision ocean monitoring.

[0291] In contrast, the method proposed in the present invention adopts a systematic multi-step processing strategy. First, various error sources and their characteristics are accurately identified through a variety of mathematical analysis methods. Then, an error correlation matrix is ​​constructed to quantify the mutual influence between errors. The error compensation weights are scientifically determined based on the entropy weight method. By designing compensation schemes for different error types, such as using polynomial compensation for instrument errors, multivariate regression compensation for correlation errors, Kalman filtering for drift errors, and deep learning models for environmental interference errors, a complete compensation system is formed. In particular, the error compensation neural network model CTSD-NET is introduced, which can effectively handle complex nonlinear error patterns that are difficult to capture with traditional methods, providing strong support for high-precision ocean observations and significantly improving the accuracy and reliability of ocean environmental parameter measurements.

[0292] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 5 and 6 below.

[0293] Table 5 Variable Explanation Table (Part I)

[0294]

[0295] Table 6 Variable Explanation Table (Part II)

[0296]

[0297] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A data processing method for a self-contained temperature, salinity and depth meter without disassembly, characterized in that: include: Collecting the raw data of the salt depth meter and performing sampling and filtering processing; Analyze the errors in the original data and identify instrument errors, correlation errors, drift errors, and environmental interference errors through various mathematical analysis methods. Calculate the correlation between errors through autocorrelation analysis, construct an error correlation matrix, and use principal component analysis to extract the main error influencing factors. Determine the compensation weights for each error based on the entropy weight method, and use corresponding compensation equations for different error types. Construct an error compensation neural network model, CTSD-NET, and train CTSD-NET using historical data to achieve comprehensive compensation for environmental interference errors. Randomly change the data sampling window length, repeat the error compensation steps, and obtain multiple sets of compensated data with different window lengths as validation sets. The compensation effect is quantitatively evaluated, and the neural network model parameters are optimized through cross-validation method; the verification steps are executed repeatedly, and the optimal result is selected to determine the final error compensation model to complete the entire data processing process.

2. The data processing method of the non-disassembly chamber self-contained temperature and salinity meter according to claim 1 is characterized in that: The raw data of the non-disassembly-type self-contained temperature, salt and depth instrument include water temperature value, conductivity value, pressure value, attitude angle value, internal battery voltage value, timestamp and internal reference sensor reading; the data sampling window duration is set to a first duration set according to the sensor frequency, and the first duration is 60 to 120 seconds.

3. The data processing method of the non-disassembly chamber self-contained temperature and salt depth instrument according to claim 2 is characterized in that: The analysis of errors in the original data specifically includes: identifying high-frequency noise through Fourier spectrum analysis to determine instrument errors, using the correlation matrix between parameters to calculate and identify correlation errors, using time series trend analysis to detect drift errors, separating environmental interference errors through wavelet transform, and identifying abnormal data points based on the adaptive threshold algorithm.

4. The data processing method of the non-disassembly chamber self-contained temperature and salt depth instrument according to claim 3 is characterized in that: The error compensation weights are determined based on the entropy weight method, and corresponding compensation equations are used for corrections according to different error types, specifically including: using a polynomial compensation equation to correct the instrument error, using a linear regression compensation equation to correct the correlation error, using a Kalman filter compensation equation to correct the drift error in real time, and using a temperature gradient compensation equation to correct the measurement error caused by temperature change.

5. The data processing method of the non-disassembly chamber self-contained temperature and salt depth instrument according to claim 4 is characterized in that: Constructing an error compensation neural network model CTSD-NET, training the neural network model through historical data to achieve comprehensive compensation for environmental interference errors specifically includes: constructing an error compensation neural network model CTSD-NET, training the error compensation neural network model CTSD-NET through historical data to achieve comprehensive compensation for the environmental interference errors, and using an error compensation equation group to optimize the compensation effect.

6. The data processing method of the non-disassembly chamber self-contained temperature and salt depth instrument according to claim 5, characterized in that: The error compensation equation group includes an instrument error compensation equation, an associated error compensation equation, a drift error compensation equation and a temperature gradient compensation equation; the instrument error compensation equation is used to correct the system deviation caused by the inherent accuracy limitation of the equipment; the associated error compensation equation is used to eliminate the measurement deviation caused by the mutual influence of various parameters; the drift error compensation equation is used to correct the accumulated system deviation during long-term measurement; the temperature gradient compensation equation is used to correct the influence of temperature changes in different water layers on measurement accuracy.

7. The data processing method of the non-disassembly chamber self-contained temperature and salt depth instrument according to claim 6, characterized in that: The specific structure of the error compensation neural network model CTSD-NET is a deep neural network architecture that combines a multi-layer bidirectional long short-term memory network with a residual connection, and includes an end-to-end learning framework consisting of a data pre-processing layer, a feature extraction layer, an error classification layer, a compensation calculation layer, and a post-processing layer.

8. The data processing method of the non-disassembly chamber self-contained temperature and salt depth instrument according to claim 7, characterized in that: The feature extraction layer adopts a multi-head attention mechanism to process the complex relationship between different types of errors. The number of attention heads of the multi-head attention mechanism is dynamically adjusted according to the number of error classifications, the number of residual connections is determined according to the length of the data sampling window, and the number of hidden layer neurons is adaptively set according to the sensor accuracy level.

9. The data processing method of the non-disassembly chamber self-contained temperature and salt depth instrument according to claim 8, characterized in that: The steps for establishing a training dataset during the pre-training process of the error-compensating neural network model CTSD-NET specifically include: selecting high-quality data samples with complete labels from historical observation data, constructing a simulated dataset with known error types and magnitudes by artificially adding error factors of varying degrees, and using calibration equipment to collect reference data under a standard environment as true value labels.

10. The data processing method of the non-disassembly chamber self-contained temperature and salt depth instrument according to claim 9, characterized in that: The training data set is divided into a training set, a validation set, and a test set in a ratio of 7:2:1, and the data is standardized to eliminate the dimensional differences between different physical quantities.

Images