Acquired data error compensation method of small-sized thermohaline instrument

Abstract

The invention provides an acquired data error compensation method for a small-sized thermohaline instrument, belongs to the technical field of electrical digital data processing, and realizes accurate error compensation by constructing a multi-stage neural network tree structure. According to the method, firstly, errors are classified into system errors, random errors and environment interference errors, and second-level and third-level subdivision is carried out to form a complete error classification system; based on the classification system, an optimal path selection function is adopted to dynamically determine a compensation strategy, and corresponding models are selected from a pre-trained miniature neural network library to form a compensation chain. All compensation model structures are subjected to optimization design, different network structures and parameter configurations are adopted for linear errors, nonlinear errors, noise errors, fluctuation errors, temperature errors and salinity mutual interference errors, the total parameter number is controlled to be at a low level, and efficient operation on small equipment is ensured. The technical problem that in the prior art, a small-sized salt thermometer is insufficient in measurement precision under the condition of low energy consumption is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of digital data processing, and particularly relates to a data collection error compensation method for a small-sized temperature and salinity instrument. BACKGROUND

[0002] The small-sized temperature and salinity instrument is an instrument and equipment widely used in the field of marine monitoring, and is used for measuring water temperature and salinity parameters. Traditional technology mainly relies on hardware calibration and a simple linear compensation method to realize measurement precision control. Such an instrument is usually deployed on a marine buoy, an underwater glider and an autonomous underwater vehicle, and needs to work stably for a long time in a low-power state.

[0003] However, the traditional temperature and salinity instrument error compensation technology has various limitations. First, the simple linear compensation method cannot deal with system nonlinear errors, especially the nonlinear characteristics exhibited by the sensor hysteresis and saturation region. Second, fixed compensation parameters cannot adapt to changing marine environments, and environmental temperature gradients and ion interference factors cause measurement deviations. Third, random errors such as high-frequency noise and random jumps cannot be solved by static calibration and need to be dynamically identified and processed.

[0004] Especially under low-energy conditions, the computing resources are limited, and complex error compensation algorithms cannot be used. The existing technology is difficult to achieve a balance between high precision and low energy consumption, which leads to insufficient measurement precision of the small-sized temperature and salinity instrument in actual application, especially the error accumulation problem when deployed for a long time. SUMMARY

[0005] Therefore, the application provides a data collection error compensation method for a small-sized temperature and salinity instrument, which can solve the technical problem of insufficient measurement precision of the small-sized temperature and salinity instrument under low-energy conditions in the prior art.

[0006] The application is implemented in the following manner. The application provides a data collection error compensation method for a small-sized temperature and salinity instrument, which comprises the following steps: constructing a multi-level neural network tree structure, each node being set as a pre-trained micro neural network, and the number of tree paths being determined according to the number of error types; collecting measurement data of the small-sized temperature and salinity instrument and comparing the measurement data with standard data to calculate error values and establish an error feature database; performing one-level classification according to error features; performing two-level classification and three-level classification on system error values, random error values and environmental interference error values, respectively; calling an optimal path selection function, inputting measurement data of the small-sized temperature and salinity instrument, environmental parameter values, historical error statistical data, device state information and measurement mode identifiers, and outputting an optimal neural network compensation path; selecting a corresponding error compensation model from a pre-trained micro neural network library according to the optimal neural network compensation path, and connecting multiple micro neural networks to form a compensation chain; and applying the compensation chain to perform multi-step error correction on the measurement data of the small-sized temperature and salinity instrument, and outputting corrected temperature and salinity values.

[0007] Wherein, the first-level classification according to error characteristics is specifically to divide error values into systematic error values, random error values and environmental interference error values; the second-level classification of systematic error values includes linear error values and nonlinear error values; the second-level classification of random error values includes noise error values and fluctuation error values; the second-level classification of environmental interference error values includes temperature error values and salinity interference error values.

[0008] Wherein, the third-level classification of linear error values includes zero-point drift error values and range offset error values; the third-level classification of nonlinear error values includes hysteresis error values and saturation error values; the third-level classification of noise error values includes high-frequency noise error values and impulse noise error values; the third-level classification of fluctuation error values includes periodic fluctuation error values and random jump error values; the third-level classification of temperature error values includes temperature gradient error values and temperature compensation deficiency error values; the third-level classification of salinity interference error values includes ion interference error values and conductivity nonlinearity error values.

[0009] Wherein, the selected error compensation model includes a systematic linear error compensation model, a systematic nonlinear error compensation model, a random noise error compensation model, a random fluctuation error compensation model, an environmental temperature error compensation model and an environmental salinity interference error compensation model.

[0010] Wherein, the specific structure of the systematic linear error compensation model is a three-layer feedforward neural network, the input layer has two neurons receiving temperature value and salinity value original measurement values, the hidden layer has four neurons adopting a Sigmoid activation function, and the output layer has two neurons respectively outputting zero-point drift error value compensation and range offset error value compensation; the total parameter quantity of the systematic linear error compensation model is controlled to be within 50.

[0011] Wherein, the specific structure of the systematic nonlinear error compensation model is a four-layer feedforward neural network, the input layer has three neurons receiving temperature value, salinity value measurement values and rate values, the hidden layer respectively contains six and four neurons, both of which adopt a ReLU activation function, and the output layer has two neurons respectively outputting hysteresis error value compensation and saturation error value compensation; the systematic nonlinear error compensation model adopts a sparse connection structure, and the total parameter quantity is controlled to be within 80.

[0012] Wherein, the specific structure of the random noise error compensation model is a recursive neural network based on wavelet transform, the input layer receives continuous five temperature value and salinity value measurement data points, contains a denoising layer and a signal reconstruction layer, has a total of three layers of structures, and the parameter quantity is controlled to be within 60; the random noise error compensation model has the ability to filter high-frequency noise error values and identify impulse noise error values.

[0013] The specific structure of the random fluctuation error compensation model is a gated recurrent unit network, the input layer receives 10 consecutive temperature value and salinity value measurement data points, the hidden layer includes 3 GRU units, and the output layer has 2 neurons respectively outputting a periodic fluctuation error value compensation amount and a random jump error value detection threshold; the total parameter amount of the random fluctuation error compensation model is controlled to be less than 70.

[0014] The specific structure of the environmental temperature error compensation model is a three-layer feedforward neural network, the input layer receives multi-point temperature value measurement values and gradient information values, the hidden layer has 5 neurons adopting a hyperbolic tangent activation function, and the output layer outputs a temperature gradient error value compensation amount and a temperature compensation deficiency error value correction amount; the total parameter amount of the environmental temperature error compensation model is controlled to be less than 60.

[0015] The specific structure of the environmental salinity mutual interference error compensation model is a four-layer feedforward neural network, the input layer receives salinity value measurement values, conductivity values and various ion concentration estimated values, two hidden layers respectively include 6 and 4 neurons, and both adopt a Leaky ReLU activation function; the output layer outputs an ion interference error value compensation amount and a conductivity nonlinear error value correction amount; the total parameter amount of the environmental salinity mutual interference error compensation model is controlled to be less than 90.

[0016] Compared with the prior art, the data collection error compensation method of the small-sized temperature and salinity instrument provided by the application can effectively solve the problem of insufficient measurement accuracy under low energy consumption conditions.

[0017] Compared with the traditional linear compensation method, the application can adapt to different working states within the full range by respectively processing zero drift, range deviation and hysteresis, saturation and other nonlinear errors through the system linear error compensation model and the system nonlinear error compensation model.

[0018] The error compensation method of the application adopts a modular micro neural network structure, each network parameter amount is controlled at a low level, and the compensation strategy is dynamically selected through an optimal path selection function, so that high-precision measurement is realized while ensuring low energy consumption. The systematic error classification and targeted compensation scheme solve the core problem of insufficient measurement accuracy of the small-sized temperature and salinity instrument under low energy consumption conditions. BRIEF DESCRIPTION OF DRAWINGS

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

[0020] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0021] As Figure 1 shown is a flow chart of a data collection error compensation method of a small temperature and salinity instrument provided by the present application, and the method comprises the following steps:

[0022] S01, a multi-level neural network tree structure is constructed, each node is set as a pre-trained micro neural network, and the number of tree paths is determined according to the number of error types;

[0023] S02, measurement data of the small temperature and salinity instrument are collected, and compared with standard data, error values are calculated, and an error feature database is established;

[0024] S03, according to the error features, a first classification is performed, and the error values are divided into three categories of system error values, random error values and environmental interference error values;

[0025] S04, a second classification is performed on the system error values, the system error values are divided into linear error values and nonlinear error values, a second classification is performed on the random error values, the random error values are divided into noise error values and fluctuation error values, and a second classification is performed on the environmental interference error values, the environmental interference error values are divided into temperature error values and salinity mutual interference error values;

[0026] S05, a third classification is performed on the linear error values, the linear error values are subdivided into zero drift error values and range offset error values, a third classification is performed on the nonlinear error values, the nonlinear error values are subdivided into hysteresis error values and saturation error values, a third classification is performed on the noise error values, the noise error values are subdivided into high-frequency noise error values and pulse noise error values, a third classification is performed on the fluctuation error values, the fluctuation error values are subdivided into periodic fluctuation error values and random jump error values, a third classification is performed on the temperature error values, the temperature error values are subdivided into temperature gradient error values and temperature compensation deficiency error values, and a third classification is performed on the salinity mutual interference error values, the salinity mutual interference error values are subdivided into ion interference error values and conductivity nonlinear error values;

[0027] S06, calling the optimal path selection function, inputting the small-sized CTD measurement data, environmental parameter values, historical error statistical data, device state information and measurement mode identification, and outputting the optimal neural network compensation path;

[0028] S07, according to the optimal neural network compensation path, selecting the corresponding system linear error compensation model, system nonlinear error compensation model, random noise error compensation model, random fluctuation error compensation model, environmental temperature error compensation model and environmental salinity mutual interference error compensation model from the pre-trained micro neural network library, and connecting multiple micro neural networks in series to form a compensation chain;

[0029] S08, applying the compensation chain to multi-step error correction of the small-sized CTD measurement data, outputting the corrected temperature and salinity values, reducing the cumulative error, and improving the measurement accuracy of the small-sized CTD under low energy consumption conditions;

[0030] S09, optionally, further comprising periodically updating the error characteristic database to ensure that the optimal path selection function can adjust the optimal neural network compensation path according to the latest data in the error characteristic database.

[0031] The micro neural network refers to a simplified neural network structure composed of a small number of neurons and having low computing resource requirements, and is suitable for running in a small device; the error feature database comprises a structured data set of temperature values, salinity values, measured environmental parameter values and corresponding error values, and is used for path selection function analysis; the system error value refers to a stability deviation value caused by the structure of the instrument itself and the characteristics of the sensor; the random error value refers to a fluctuation deviation value caused by uncertain factors in the measurement process; the environmental interference error value refers to a measurement deviation value caused by changes in external environmental conditions; the neural network path refers to a complete connection path from a root node to a leaf node of a neural network tree, and represents an error compensation strategy; the compensation chain refers to a data processing sequence formed by connecting a plurality of micro neural networks in a specific order, and each network processes a type of error value; the zero drift error value refers to a deviation value generated by the change of the output value of the sensor in the zero input state over time; the range offset error value refers to a deviation value caused by the difference in the slope of the sensor output and the ideal linear relationship; the hysteresis error value refers to a deviation value caused by the inconsistency of the output value in the process of increasing and decreasing the input of the sensor; the saturation error value refers to a deviation value caused by the nonlinear growth of the sensor response when approaching the range limit; the high-frequency noise error value refers to a high-frequency random signal superimposed deviation value caused by circuit interference; the impulse noise error value refers to a large amplitude deviation value caused by a strong burst interference signal in a short time; the periodic fluctuation error value refers to a cyclic measurement deviation value with fixed frequency characteristics; the random jump error value refers to a deviation value caused by irregular burst measurement value abnormal change; the temperature gradient error value refers to a deviation value caused by the non-uniform temperature of different parts inside the sensor; the temperature compensation deficiency error value refers to a residual deviation value caused by the failure of the existing temperature compensation mechanism to completely eliminate the temperature influence; the ion interference error value refers to a deviation value caused by the interference of non-target ions in the water body to the salinity measurement; the conductivity nonlinearity error value refers to a deviation value caused by the nonlinear relationship between conductivity and concentration in a high-concentration salt solution; the environmental parameter value refers to the temperature value, pressure value, humidity value and water flow rate value of the measurement environment; the historical error statistical data refers to the statistical distribution and change trend of the error values recorded in the past measurement process; the device state information refers to the battery capacity value, sensor usage time value and last calibration time value of the small temperature and salinity instrument; and the measurement mode identifier refers to the working mode of the small temperature and salinity instrument, including a static measurement mode identifier, a dynamic measurement mode identifier, a continuous monitoring mode identifier and a fixed-point sampling mode identifier.

[0032] The optimal path selection function is used to determine the most suitable neural network compensation path according to the current measured environment and data characteristics, and the input includes the small-sized thermosalinograph measurement data, the environmental parameter value, the historical error statistical data, the device state information and the measurement mode identifier, and the output is the optimal neural network compensation path; the optimal path selection function comprehensively evaluates the compensation effect of each neural network path by analyzing the small-sized thermosalinograph measurement data characteristics, the environmental parameter value similarity, the historical error statistical data mode matching degree and the device state information influence factor, selects the neural network combination with the strongest error compensation ability, and forms the optimal neural network compensation path.

[0033] The specific structure of the system linear error compensation model is a three-layer feedforward neural network, the input layer has two neurons receiving the temperature value and the salinity value original measurement value, the hidden layer has four neurons adopting a Sigmoid activation function, and the output layer has two neurons respectively outputting the zero drift error value compensation amount and the range offset error value compensation amount; the total parameter amount of the system linear error compensation model is controlled to be less than 50, so as to ensure efficient operation on the small-sized thermosalinograph.

[0034] The specific structure of the system nonlinear error compensation model is a four-layer feedforward neural network, the input layer has three neurons receiving the temperature value, the salinity value measurement value and the rate value, the hidden layer respectively includes six and four neurons, and both adopt a ReLU activation function, and the output layer has two neurons respectively outputting the hysteresis error value compensation amount and the saturation error value compensation amount; the system nonlinear error compensation model adopts a sparse connection structure, and the total parameter amount is controlled to be less than 80.

[0035] The specific structure of the random noise error compensation model is a recursive neural network based on wavelet transform, the input layer receives continuous five temperature value and salinity value measurement data points, includes a denoising layer and a signal reconstruction layer, and has a total of three layers of structures, and the parameter amount is controlled to be less than 60; the random noise error compensation model has the ability of filtering the high-frequency noise error value and identifying the pulse noise error value.

[0036] The specific structure of the random fluctuation error compensation model is a gated recurrent unit network, the input layer receives continuous ten temperature value and salinity value measurement data points, the hidden layer includes three GRU units, and the output layer has two neurons respectively outputting the periodic fluctuation error value compensation amount and the random jump error value detection threshold; the total parameter amount of the random fluctuation error compensation model is controlled to be less than 70.

[0037] The specific structure of the ambient temperature error compensation model is a three-layer feedforward neural network, the input layer receives the temperature value measurement and its gradient information value, the hidden layer has 5 neurons using hyperbolic tangent activation function, and the output layer outputs the temperature gradient error value compensation and the temperature compensation error value correction; the total parameter quantity of the ambient temperature error compensation model is controlled within 60.

[0038] The specific structure of the ambient salinity mutual interference error compensation model is a four-layer feedforward neural network, the input layer receives the salinity value measurement, conductivity value and various ion concentration estimation value, two hidden layers respectively contain 6 and 4 neurons, both using Leaky ReLU activation function, and the output layer outputs the ion interference error value compensation and the conductivity nonlinear error value correction; the total parameter quantity of the ambient salinity mutual interference error compensation model is controlled within 90.

[0039] The steps of establishing the training data set in the pre-training process of the system linear error compensation model specifically include measuring different concentration salt solution sample data using high-precision reference instrument and the small-sized temperature salinity instrument under standard laboratory environment, recording the measurement difference of the two under different temperature value conditions, extracting the linear error value characteristics in the system error value and marking the zero drift error value and the range offset error value components, and constructing a training data set containing 1000 groups of samples, each group of samples containing original temperature value and salinity value measurement, standard temperature value and standard salinity value, and zero drift error value compensation and range offset error value compensation target value.

[0040] The steps of establishing the training data set in the pre-training process of the system nonlinear error compensation model specifically include designing dynamic measurement experiment, making the small-sized temperature salinity instrument rise and fall to measure the same sample within the full range, recording the hysteresis error value data; at the same time, high-density sampling is carried out near the range boundary area to capture the nonlinear characteristics of the saturation error value, and a training set containing 1500 groups of samples is constructed, covering various nonlinear error value conditions and working conditions.

[0041] The steps of establishing the training data set in the pre-training process of the random noise error compensation model specifically include collecting the temperature value and salinity value data under various electromagnetic interference environments, introducing the high-frequency noise error value and the pulse noise error value, and recording the reference data under the interference-free condition as the target value, and constructing a training set containing 2000 groups of time sequence samples, each sequence containing 5 consecutive temperature value and salinity value data points and corresponding denoising target temperature value and denoising target salinity value.

[0042] The step of establishing the training data set in the random fluctuation error compensation model pre-training process specifically includes continuously collecting the temperature value and salinity value data for a long time, identifying and marking the periodic fluctuation error value mode and the random jump error value event, extracting time sequence features, and constructing a training set containing 500 long sequence samples, each containing 10 consecutive temperature value and salinity value measurement points and corresponding periodic fluctuation error value prediction and random jump error value detection labels.

[0043] The step of establishing the training data set in the ambient temperature error compensation model pre-training process specifically includes placing a multi-point temperature sensor and the small-sized temperature and salinity instrument in a temperature gradient control environment, creating various temperature gradient conditions and recording measurement deviations, while evaluating the effect of existing temperature compensation mechanisms, constructing a training set containing 1200 multi-temperature point samples, covering various temperature gradient error value modes and temperature compensation insufficient error value conditions.

[0044] The step of establishing the training data set in the ambient salinity mutual interference error compensation model pre-training process specifically includes preparing salt solution samples containing different concentrations of interfering ions, recording the relationship between the conductivity value and the actual salinity value, measuring the conductivity nonlinear error value response characteristics under high concentration conditions, and constructing a training set containing 800 multi-parameter samples, each containing the conductivity value, the ion concentration estimated value, and corresponding ion interference error value compensation target value and conductivity nonlinear error value correction target value.

[0045] The step of pre-training the system linear error compensation model specifically includes applying a small batch gradient descent algorithm, setting the batch size to 16, setting the initial learning rate to 0.01 and using the cosine annealing strategy, selecting the mean square error as the loss function, training for 200 rounds, using early stopping technology to avoid overfitting, and finally verifying that the error of the final validation set is less than 0.05% full scale.

[0046] The step of pre-training the system nonlinear error compensation model specifically includes applying the Adam optimizer, setting the initial learning rate to 0.005, setting the batch size to 32, combining the mean square error and the smooth L1 loss as the loss function, training for 300 rounds, using weight decay regularization to reduce the risk of overfitting, and finally the system nonlinear error compensation model has a hysteresis error compensation accuracy better than 0.08% full scale.

[0047] The step of pre-training the random noise error compensation model specifically includes applying the momentum optimizer, setting the learning rate to 0.008, setting the batch size to 64, combining the signal reconstruction error and the frequency domain feature preservation term as the loss function, training for 400 rounds, using data augmentation technology to increase the robustness of the random noise error compensation model, and finally the random noise error compensation model can effectively remove more than 90% of the high-frequency noise error value components.

[0048] The step of pre-training the random fluctuation error compensation model specifically includes applying an RMSprop optimizer with a learning rate of 0.003, a batch size of 8, using a truncated backpropagation algorithm for long sequence characteristics, a loss function combining prediction error and random jump error value detection F1 score, training for 250 rounds, and finally the random fluctuation error compensation model can accurately predict the trend of the periodic fluctuation error value and detect more than 95% of the random jump error value.

[0049] The step of pre-training the ambient temperature error compensation model specifically includes applying an AdamW optimizer with a learning rate of 0.006, a batch size of 24, a loss function setting weights for temperature gradient error value compensation and temperature compensation deficiency error value correction, training for 350 rounds, and using a learning rate scheduler. Finally, the ambient temperature error compensation model has a compensation accuracy better than 0.07% full scale under various temperature gradient conditions.

[0050] The step of pre-training the ambient salinity mutual interference error compensation model specifically includes applying a Nadam optimizer with a learning rate of 0.004, a batch size of 16, a loss function combining ion interference error value compensation error and conductivity nonlinearity error value correction accuracy, training for 320 rounds, and using a hierarchical learning rate strategy to accelerate convergence. Finally, the ambient salinity mutual interference error compensation model has a compensation accuracy better than 0.1% full scale under high salinity conditions.

[0051] The specific implementation of the above steps is described in detail below.

[0052] The specific implementation of step S01 is to construct a hierarchical nested multi-level neural network tree structure composed of pre-trained micro neural network nodes. First, the root node of the tree is designed as an error classifier, responsible for preliminary identification of error types. Then, 3 nodes are set in the second layer corresponding to system error, random error and environmental interference error. Six nodes are set in the third layer, corresponding to linear error, non-linear error, noise error, fluctuation error, temperature error and salinity mutual interference error. Twelve leaf nodes are set in the fourth layer, corresponding to zero drift error, range offset error, hysteresis error, saturation error, high-frequency noise error, impulse noise error, periodic fluctuation error, random jump error, temperature gradient error, temperature compensation deficiency error, ion interference error and conductivity nonlinearity error. The nodes are connected by selective connection to form a tree-shaped compensation path, each path representing a specific error correction strategy. The depth and width of the tree are dynamically adjusted according to the complexity and number of error types to ensure compensation accuracy while minimizing computational overhead. This tree structure uses the hierarchical idea to decompose the complex error compensation problem into multiple simple sub-problems, realizes modular error processing, and improves system adaptability and maintainability.

[0053] The specific implementation of step S02 is to establish a complete error feature database. First, the small-scale temperature and salinity instrument and the high-precision standard equipment are simultaneously placed in a variety of controlled environments (including temperature in the range of 0°C to 40°C, salinity in the range of 0 psu to 40 psu) for synchronous measurement. During the collection process, the environmental parameters are kept stable for 300 seconds, and at least 50 data points are obtained in each stable period. For each set of measurement values, the temperature difference and salinity difference between the small-scale temperature and salinity instrument and the standard equipment are calculated as the original error data. Then, combined with the time stamp, environmental parameters and equipment state information, a multi-dimensional error feature vector is constructed. Principal component analysis is used for dimension reduction to extract the key factors affecting the error. Finally, the processed error data is indexed and classified according to time sequence, environmental conditions and error amplitude to form a structured error feature database. This database not only contains static error data, but also records the dynamic characteristics of error changes over time and environment, providing a data basis for subsequent error classification and compensation path selection.

[0054] The specific implementation of step S03 is to perform a primary classification of errors. First, the K-means clustering algorithm is used to preliminarily cluster the error feature data, and the number of clusters is set to 3, representing three main error types. Then, based on time stability indicators (such as identifying system errors when the variance changes less than 0.05% over time), spectral characteristics (such as identifying random errors when high-frequency components are included and the signal-to-noise ratio is less than 15 dB), and environmental correlation coefficients (such as identifying environmental interference errors when the absolute value of the correlation coefficient with temperature or salinity gradient is greater than 0.7), the error values are classified into three categories: system error values, random error values, and environmental interference error values. For boundary cases that are difficult to classify directly, fuzzy logic rules are used for judgment, and a membership function is introduced to calculate the probability of error belonging to each type, and the type corresponding to the highest probability is taken as the classification result. This classification process uses the essential characteristics of errors for classification, laying the foundation for subsequent more detailed classification and targeted compensation.

[0055] The specific implementation of step S04 is to perform a two-level classification of errors. For systematic error values, classification is performed by a linear correlation test, the Pearson correlation coefficient of the error and the measured value is calculated, and when the absolute value of the correlation coefficient is greater than 0.85 and the error changes with the measured value in a clear linear relationship, it is classified as a linear error value; when the absolute value of the correlation coefficient is less than 0.85 or the error and the measured value show a clear nonlinear characteristic, it is classified as a nonlinear error value. For random error values, classification is performed by frequency domain analysis, and the frequency spectrum characteristics of the error are extracted by applying fast Fourier transform, and when the frequency spectrum energy is mainly concentrated in the high frequency band (more than 1 / 4 of the sampling frequency), it is classified as a noise error value; when the frequency spectrum energy is widely distributed and there is a clear low-frequency fluctuation characteristic, it is classified as a fluctuation error value. For environmental interference error values, classification is performed by environmental sensitivity analysis, and the partial derivative of the error and the temperature change is calculated, and when the temperature sensitivity exceeds 0.1% / °C of the full scale, it is classified as a temperature error value; when the error and the salinity measured value or the conductivity exist cross-influence relationship, it is classified as a salinity mutual interference error value. This two-level classification method uses the mathematical characteristics and physical mechanisms of the error to perform more accurate classification, providing direction for subsequent fine compensation.

[0056] The specific implementation of step S05 is to perform three-level classification of errors. For linear error values, by analyzing the linear regression equation of the error and the measured value, when the absolute value of the intercept is greater than 0.3% of the full scale, it is classified as a zero drift error value; when the slope deviates from the ideal value by more than ±5%, it is classified as a range offset error value. For non-linear error values, by comparing rising and falling measurements, when the error difference of the same measurement point in the rising process and the falling process is greater than 0.2% of the full scale, it is classified as a hysteresis error value; when the error shows an accelerated growth trend near the range boundary and the derivative is greater than 2 times the normal area, it is classified as a saturation error value. For noise error values, by time-frequency analysis, when the error contains a persistent high-frequency component and the frequency is higher than 1 / 3 of the measurement frequency, it is classified as a high-frequency noise error value; when there are abnormal points in the error with an amplitude greater than 3 times the standard deviation and a duration shorter than 3 sampling periods, it is classified as an impulse noise error value. For fluctuation error values, by autocorrelation analysis, when the autocorrelation function of the error presents obvious periodicity and the period is stable, it is classified as a periodic fluctuation error value; when the error has irregular mutations and the change amplitude is greater than 2 times the average fluctuation amplitude, it is classified as a random jump error value. For temperature error values, by multi-point temperature measurement analysis, when the correlation coefficient of the error and the temperature gradient (temperature spatial change rate) is greater than 0.6, it is classified as a temperature gradient error value; when the residual error after temperature compensation still shows correlation with temperature changes, it is classified as an insufficient temperature compensation error value. For salinity mutual interference error values, by ion selectivity test, when the error changes with the change of the concentration of a specific ion, it is classified as an ion interference error value; when the error shows nonlinear growth under high salinity conditions (greater than 30 psu), it is classified as a conductivity nonlinear error value. This three-level classification process uses detailed characteristics of errors for accurate classification, laying a foundation for subsequent accurate compensation.

[0057] The specific implementation of step S06 is to call the optimal path selection function to determine the best compensation strategy. The function uses a decision fusion algorithm based on evidence theory to integrate multi-source information to determine the optimal compensation path. First, singular spectrum analysis is applied to the input small-scale salinity instrument measurement data to extract the main mode characteristics and abnormal characteristics of the data, and the similarity between the feature vector and each type of error feature is calculated. Then, an environmental parameter similarity matrix is constructed, and the similarity between the current environmental parameters and the environmental parameters in the historical data is measured by Euclidean distance, with a threshold of 0.15. Next, pattern matching of historical error statistical data is performed, and the dynamic time warping algorithm is used to calculate the matching degree of the current error pattern and the historical error pattern, with a threshold of 0.75. At the same time, the influence factors are calculated according to the device state information, including the battery power attenuation coefficient (the weight increases when the power is less than 30%), the sensor aging coefficient (the weight increases when the service time exceeds 500 hours) and the calibration time coefficient (the weight increases when the distance to the last calibration exceeds 30 days). Finally, consider the measurement mode identifier, and assign different compensation strategy weights to different modes (static, dynamic, continuous monitoring, and fixed-point sampling). Based on the above analysis results, the analytic hierarchy process is used to calculate the comprehensive score of each candidate compensation path, and the path with the highest score is selected as the optimal neural network compensation path. This method considers multiple dimensions and can adaptively select the most suitable compensation strategy for the current measurement conditions, improving the accuracy and stability of the compensation effect.

[0058] The specific implementation of step S07 is to construct a compensation chain according to the optimal neural network compensation path. First, extract the neural network model corresponding to the optimal path from the pre-trained micro neural network library. This process uses a model retrieval algorithm to quickly locate the required model file based on the path identifier. Then, determine the model series order according to the error propagation law, usually following the basic order of system error compensation → random error compensation → environmental error compensation, to ensure that the compensation of the previous stage does not have a negative impact on the next stage. For system linear error compensation models and system nonlinear error compensation models, they can run in parallel when there is no obvious dependence between them to improve efficiency. For random noise error compensation models and random fluctuation error compensation models, the former handles high-frequency characteristics and the latter handles low-frequency characteristics, so the noise model is usually applied first and then the fluctuation model. For environmental temperature error compensation models and environmental salinity mutual interference error compensation models, considering that temperature usually affects salinity measurement, the temperature model is applied first and then the salinity model. Design a data interface between models to ensure consistent data format conversion, and the output of each model is standardized before being transmitted to the next model. Finally, apply model fusion technology and set the weight coefficients of each model to dynamically adjust the contribution of each model to the final compensation result according to the current error characteristics. This compensation chain construction method based on the optimal path fully utilizes the modular design concept and realizes flexible and efficient error compensation.

[0059] The specific implementation of step S08 is to perform multi-step error correction by applying a compensation chain. First, the original measurement data of the small-scale temperature and salinity instrument and the current environmental parameters are input as inputs into the first model of the compensation chain. Each model receives the output of the previous model (or the original input) for processing, calculates a specific type of error compensation amount, and passes the corrected data to the next model. For the system linear error compensation model, the zero drift error compensation amount and the range offset error compensation amount are calculated by feedforward, corresponding to the addition correction and multiplication correction, respectively. For the system nonlinear error compensation model, the input data and its rate of change are processed, and the hysteresis error compensation amount and the saturation error compensation amount are output to correct the measurement deviation in the nonlinear region. For the random noise error compensation model, the wavelet transform denoising algorithm is applied to the continuous data points to filter out high-frequency noise and detect pulse interference. For the random fluctuation error compensation model, the GRU structure is used to capture the time series characteristics, predict and eliminate periodic fluctuations, and mark abnormal jump points. For the environmental temperature error compensation model, the temperature-related compensation amount is calculated based on the temperature gradient information at multiple points. For the environmental salinity mutual interference error compensation model, ion interference and high salinity nonlinear effects are processed. After all the models are processed in series, the final corrected temperature and salinity values are output. Throughout the process, an error propagation control strategy is adopted to ensure that each level of compensation is coordinated with each other, avoiding error accumulation and overcompensation. This method greatly improves the measurement accuracy through multi-level fine compensation, enabling the small-scale temperature and salinity instrument to achieve measurement performance close to high-end equipment under low-power consumption conditions.

[0060] Step S09 is an optional step, and its specific implementation is to periodically update the error feature database to adapt to the dynamic changes of the device and the environment. First, an automatic data collection mechanism is established to store measurement data, environmental parameters, and corresponding error analysis results periodically (e.g., every 100 measurements) during normal measurement. Then, a data screening algorithm is designed to remove unreliable data through anomaly detection techniques (such as isolation forest method, with an anomaly threshold of 1.5% full scale) and retain representative error samples. Next, an incremental learning strategy is implemented to integrate new data with the existing database, and the error feature statistical parameters are updated using a weighted average method, with a new data weight coefficient of 0.3. At the same time, a time decay mechanism is applied to gradually reduce the influence of old data over time, with a decay coefficient of 0.05 / month. When the cumulative amount of new data exceeds 10% of the database capacity or a significant change in error features is detected (e.g., a change in the mean value of a certain type of error exceeding 30%), the database reconstruction process is triggered to reorganize the index structure and optimize storage. Finally, the updated error feature database is automatically synchronized to the optimal path selection function to ensure that path selection is always based on the latest error features. This implementation adopts an incremental database maintenance strategy, enabling dynamic updating and adaptive evolution of the error feature library, allowing the compensation system to respond to new error patterns caused by device aging, environmental changes, and other factors in a timely manner, maintaining long-term compensation effectiveness.

[0061] The mathematical model or calculation process involved in the present application is described in detail below.

[0062] In step S03, the K-means clustering algorithm is used to preliminarily cluster the error feature data, and the calculation process can be represented as:

[0063] In the formula, J is the clustering objective function; E i is the feature vector of the i-th error sample; C j is the j-th cluster center; and N is the total number of error samples. The equation is based on the Euclidean distance metric, and the error samples are divided into three categories by minimizing the sum of the squares of the distances from the error samples to the cluster centers.

[0064] Calculation of the time stability index of the error:

[0065] In the formula, S t is the time stability index; σ 2 (t) is the variance of the error at time t; σ 2 (t+Δt) is the variance of the error at time t+ΔT; and ΔT is the time interval. When S t <0.05%, it is identified as a systematic error. The equation calculates the relative change rate of the error variance to judge the stability of the error in the time dimension.

[0066] Calculation of the signal-to-noise ratio of the spectral characteristics:

[0067] In the formula, SNR is the signal-to-noise ratio, with the unit of decibel (dB); P signal is the signal power; and P noise is the noise power. When SNR<15dB and contains high-frequency components, it is identified as a random error. The equation is used to evaluate the influence degree of the noise in the measured signal.

[0068] Calculation of the environmental correlation coefficient:

[0069] In the formula, ρ E，P is the correlation coefficient between the error and the environmental parameter; Cov(E, P) is the covariance between the error E and the environmental parameter P; σ E is the standard deviation of the error; and σ P is the standard deviation of the environmental parameter. When |ρ E，P |>0.7, it is identified as an environmental interference error. The equation calculates the correlation between the error and the environmental parameter to judge the influence degree of the environmental factors on the measurement.

[0070] Calculation of the membership function in fuzzy logic:

[0071] where μ j (E i ) is the error sample E i the membership degree of the jth error; ||E i -C j || is the Euclidean distance between error sample and cluster center; d0 is the standardization parameter, which is the average distance between samples; p is the fuzzy factor, which is 2. This equation is based on the fuzzy membership degree calculation principle of distance, which is used to handle the boundary situation that is difficult to classify directly.

[0072] In step S04, the Pearson correlation coefficient of linear correlation test is calculated:

[0073]

[0074] where r is the Pearson correlation coefficient; E i is the ith error value; is the error mean; M i is the ith measurement value; is the measurement mean; n is the sample number. When |r|>0.85, it is classified as a linear error value. This equation is used to evaluate the degree of linear correlation between error and measurement value.

[0075] Fast Fourier Transform is used for frequency domain analysis, and the calculation process is:

[0076] where X k is the frequency domain representation; x n is the time domain error sequence; N is the sequence length; k is the frequency index; i is the imaginary unit. This equation reveals the frequency characteristics of the error by converting the time domain signal to the frequency domain.

[0077] High frequency energy ratio calculation:

[0078] where R HF is the high frequency energy ratio; |X k | is the frequency domain amplitude; N is the FFT point number. When R HF is large, it is classified as a noise error value. This equation calculates the energy ratio of high frequency band, which is used to identify noise type error.

[0079] Environmental sensitivity calculation:

[0080] where S T is the temperature sensitivity; is the partial derivative of error with respect to temperature; ΔE is the error change; ΔT is the temperature change. When S TValues ​​greater than 0.1% of full scale per ℃ are classified as temperature error values. This equation measures the degree to which the error is affected by temperature changes.

[0081] In step S05, the linear regression equation is calculated as: E = β0 + β1M + ε;

[0082] In the formula, E is the error value; β0 is the intercept, representing zero drift; β1 is the slope, representing range offset; M is the measured value; and ε is the random error term. When |β0| > 0.3% of full scale, it is classified as a zero drift error value; when |β1-1| > 0.05, it is classified as a range offset error value. This equation simplifies the relationship between error and measured value into a linear model, facilitating the analysis of the system's linear error characteristics.

[0083] Hysteresis error calculation: E hys =|E up (M)-E down (M)|;

[0084] In the formula, E hys E represents the hysteresis error value. up (M) represents the error when the measured value is M during the ascent; E down (M) represents the error when the measured value is M during the descent. When E hys Errors greater than 0.2% of full scale are classified as hysteresis errors. This equation determines the hysteresis characteristic by comparing the error differences at the same measurement point during the rising and falling processes.

[0085] Comparison of derivatives for saturation error determination:

[0086] In the formula, R sat The ratio of derivatives; dE / dM| bound The derivative of the error with respect to the measured value near the range boundary; dE / dM| normal This is the derivative of the normal range error with respect to the measured value. When R... sat When the value is greater than 2, it is classified as a saturation error value. This equation determines the saturation characteristic by comparing the error change rate between the boundary region and the normal region.

[0087] Impulse noise determination:

[0088]

[0089] In the formula, P i For pulse indication; E i Let be the i-th error value; μ be the mean error; σ be the standard deviation of the error; d i The abnormal duration (number of sampling periods) is represented by this equation. This equation determines impulse noise based on statistical characteristics and duration.

[0090] Autocorrelation function calculation:

[0091] where R xx (k) is the autocorrelation function with a time lag of k; E i is the ith error value; μ is the mean error; and N is the sequence length. This equation reveals periodic characteristics by calculating the correlation of an error sequence with its time-shifted version.

[0092] Random jump determination:

[0093] where J i is the jump flag; E i is the ith error value; and is the average error variation. This equation determines a random jump by comparing whether the variation of adjacent error values exceeds twice the normal fluctuation.

[0094] Temperature gradient correlation coefficient calculation:

[0095] where R is the correlation coefficient of error and temperature gradient; E is the error E and temperature gradient Cov(E E , T) is the covariance of error and temperature; and σ is the standard deviation of error; and σ is the standard deviation of temperature. When

[0096] Temperature compensation deficiency error determination:

[0097] where R TC is the correlation coefficient of residual error after temperature compensation and temperature; E comp is the residual error after temperature compensation; T is the temperature value; Cov(E comp , T) is the covariance of residual error after compensation and temperature; is the standard deviation of residual error after compensation; and σ T is the standard deviation of temperature. This equation evaluates the sufficiency of existing temperature compensation by calculating the correlation of error after temperature compensation and temperature.

[0098] In step S06, the singular spectrum analysis calculation process is:

[0099] First, construct the trajectory matrix:

[0100] where X is the trajectory matrix; x iis the time series data point; L is the window length; K = N - L + 1, N is the sequence length. This matrix reconstructs one-dimensional time series into a multi-dimensional matrix by sliding window.

[0101] Then singular value decomposition is performed: X = U∑V T ;

[0102] In the formula, U is the left singular vector matrix; ∑ is the singular value diagonal matrix; V T is the transpose of the right singular vector matrix. This decomposition decomposes the trajectory matrix into principal components and noise components.

[0103] Feature vector and error feature similarity calculation:

[0104] In the formula, S(F, F j ) is the cosine similarity of feature vector F and error feature F j ; F·F j is the dot product of two vectors; ||F|| and ||F j || are the Euclidean norms of two vectors. This equation calculates the cosine similarity between vectors to determine the similarity of the current data to each error feature.

[0105] Environmental parameter similarity matrix calculation:

[0106] In the formula, D ij is the Euclidean distance between the current environmental parameter i and the historical environmental parameter j; P ik is the kth component of the current environmental parameter vector; P jk is the kth component of the historical environmental parameter vector; m is the environmental parameter dimension. This equation calculates the distance between environmental parameter vectors to evaluate environmental similarity.

[0107] Distance matrix calculation of dynamic time warping algorithm:

[0108] DTW(i, j) = d(i, j) + min{DTW(i-1, j), DTW(i, j-1), DTW(i-1, j-1)};

[0109] In the formula, DTW(i, j) is the minimum matching distance of the first i points and the first j points of the sequence; d(i, j) is the distance between the ith point and the jth point. This equation calculates the optimal alignment of two time series by dynamic programming to evaluate the matching degree of error patterns.

[0110] Device state influence factor calculation:

[0111] F device = w b ·F b +ws ·F s +w c ·F c ;

[0112] In the formula, F device F is the comprehensive influencing factor of equipment condition. b This is the battery capacity degradation coefficient, which is 2 when the battery capacity is below 30%, and 1 otherwise; F s This is the sensor aging factor, which is 1.5 when the usage time exceeds 500 hours, and 1 otherwise; F c The calibration time factor is set to 1.8 when more than 30 days have passed since the last calibration, and 1 otherwise; w b w s w c These represent the weights of the three coefficients, each defaulting to 1 / 3. This equation comprehensively considers multiple aspects of the equipment's condition to assess its impact on measurement accuracy.

[0113] Calculation of the overall score for the Analytic Hierarchy Process (AHP):

[0114] In the formula, Score i The comprehensive score for the i-th compensation path; w j C represents the weight of the j-th evaluation indicator. ij Let be the score of the i-th path on the j-th indicator; n be the number of evaluation indicators. This equation uses weighted summation to score the compensation path based on all indicators, and selects the path with the highest score as the optimal choice.

[0115] In step S08, the system linearity error compensation is calculated:

[0116] T comp1 =T raw -C zero，T +T raw ·C span，T ;

[0117] S comp1 =S raw -C zero，S +S raw ·C span，S ;

[0118] In the formula, T comp1 and S comp1 These are the temperature and salinity values ​​after system linearity error compensation, respectively; T raw and S raw The original measured temperature and salinity values; C zero，T and C zero，S These are the zero-point drift error compensation amounts for temperature and salinity, respectively; C span，T and Cspan，S These represent the range offset error compensation amounts for temperature and salinity, respectively. This equation compensates for the system's linearity error through a combination of additive / subtractive corrections and proportional corrections.

[0119] System nonlinear error compensation calculation:

[0120] T comp2 =T comp1 -C hys，T (T comp1 ,ΔT / Δt)-C sat，T (T comp1 );

[0121] S comp2 =S comp1 -C hys，S (S comp1 ,ΔS / Δt)-C sat，S (S comp1 );

[0122] In the formula, T comp2 and S comp2 These are the temperature and salinity values ​​after system nonlinear error compensation, respectively; C hys，T and C hys，S C represents the hysteresis error compensation functions for temperature and salinity, respectively, and is related to the measured values ​​and their rates of change; sat，T and C sat，S These are the saturation error compensation functions for temperature and salinity, respectively, and are related to the measured values. This equation compensates for the nonlinear errors of the system by considering the measured values ​​and their variation characteristics.

[0123] Wavelet transform calculation for random noise error compensation:

[0124] In the formula, W j，k ψ represents the wavelet coefficients; x(t) represents the signal; ψ represents the wavelet basis function; j represents the scaling parameter; and k represents the translation parameter. This equation decomposes the signal into different frequency components through multi-scale analysis, facilitating noise separation.

[0125] Wavelet thresholding denoising calculation:

[0126] In the formula, The wavelet coefficients after thresholding; W j，k λ represents the original wavelet coefficients. j The threshold corresponding to scale j is usually taken as... Where σ j Let be the noise standard deviation at scale j, and N be the signal length. This equation filters out noise components through thresholding.

[0127] GRU calculation for random fluctuation error compensation:

[0128] z t =σ(W z ·[h t-1 x t ]+b z );

[0129] r t =σ(W r ·[h t-1 x t ]+b r );

[0130]

[0131] In the formula, z t For updating the gate; r t To reset the door; For candidate hidden states; h t The current hidden state; x t For the input sequence; W z W r W is the weight matrix; b z b r σ and b are bias vectors; σ is the sigmoid activation function; and tanh is the hyperbolic tangent activation function. This series of equations implements the forward computation process of GRU, capturing long-term dependencies in time series.

[0132] Calculation of ambient temperature error compensation:

[0133]

[0134] In the formula, T comp4 and S comp4 These are the temperature and salinity values ​​after compensation for ambient temperature error; This is the error compensation amount caused by the temperature gradient, which is the temperature gradient. Functions; C insuf (T, ΔT) represents the error compensation amount caused by insufficient temperature compensation, which is related to temperature T and its change ΔT. This is the compensation amount for the effect of temperature on salinity measurement. This equation compensates for the influence of ambient temperature on the measurement by taking into account temperature and its distribution characteristics.

[0135] Calculation of environmental salinity interference error compensation:

[0136] S comp5 =S comp4 -C ion (I1, I2, ..., I n )-C nonlin (EC, S) comp4 );

[0137] where S comp5 is the compensated salinity value; C ion (I1, I2,..., I n ) is the ion interference error compensation, which is a function of ion concentrations I1, I2,..., I n ; C nonlin (EC, S comp4 ) is the conductivity nonlinearity error compensation, which is related to conductivity EC and compensated salinity value S comp4 . This equation compensates the environmental interference of salinity measurement by considering the composition and conductivity characteristics of the water body.

[0138] Final output calculation: T final = T comp4 ; S final = S comp5 ;

[0139] where T final and S final are the final temperature output value and salinity output value after the complete compensation chain processing. This equation marks the completion of multi-stage error compensation, and outputs the final corrected measurement results.

[0140] In step S09, the weighted average calculation of error feature database update:

[0141] F new = (1 - a) · F old + a · F recent ;

[0142] where F new is the updated error feature parameter; F old is the original error feature parameter; F recent is the error feature parameter of newly collected data; a is the new data weight coefficient, taking a value of 0.3. This equation balances the influence of historical data and new data through weighted average, achieving incremental update of the database.

[0143] Time decay mechanism calculation:

[0144]

[0145] where w i is the weight of the i-th data; l is the decay coefficient, taking a value of 0.05 / month; t i is the number of months from the current time. This equation makes the influence of old data decay exponentially over time, ensuring that the database reflects the latest error characteristics.

[0146] Isolation forest algorithm for anomaly detection calculates anomaly score:

[0147] In the formula, s(x, n) is the anomaly score of sample x; E(h(x)) is the average path length of sample x; c(n) is the average path length of the binary search tree with sample number n, and c(n) = 2H(n-1)-(2(n-1) / n) is usually taken, wherein H(i) is the harmonic number. When s(x, n) > 1.5% full scale, the sample is regarded as an abnormal value. The equation calculates the abnormal degree of the sample based on the isolation principle and is used for screening reliable data.

[0148] Specifically, the principle of the application is that the technical principle core of the application is to decompose the complex CTD instrument error problem into hierarchical sub-problems, and to realize directional compensation by applying resource-efficient micro neural networks. This method is based on two basic theories of error source analysis and model optimization.

[0149] Firstly, the application establishes a three-level error classification system by systematically analyzing the error sources of the CTD instrument. In the first level, the error is divided into systematic error, random error and environmental interference error; in the second level, it is further refined into linear error and nonlinear error, noise error and fluctuation error, temperature error and salinity mutual interference error; and in the third level, it is more finely divided into specific zero drift, range offset, hysteresis, saturation and other error types. This hierarchical error classification framework makes the complex error compensation problem be decomposed into several manageable sub-problems, and each sub-problem can be handled by a special micro neural network.

[0150] Secondly, the application designs micro neural networks with corresponding structures for different types of error characteristics. For example, for linear system error, a simple three-layer feedforward neural network is used; for random fluctuation error with obvious time sequence characteristics, a gated recurrent unit network is used; and for noise problems, a recursive neural network based on wavelet transform is used. These network models are lightweight, with parameter quantity controlled within a reasonable range (50-90 parameters), and are suitable for running on small devices with limited computing resources.

[0151] Thirdly, the application introduces an optimal path selection function to dynamically select the most suitable neural network compensation path according to the current measurement conditions, environmental parameters, historical error statistics and device state. This adaptive compensation strategy can maintain the best compensation effect under different working conditions and avoid the limitations of fixed compensation schemes.

[0152] Finally, by finely pre-training various types of micro neural networks, the application ensures the generalization ability and performance stability of the compensation model. The construction of the training data set covers various error conditions and working environments, efficient algorithms such as gradient descent and Adam optimizer are applied in the training process, and techniques such as early stopping and regularization are used to prevent overfitting, ensuring the effectiveness of the model in actual application.

[0153] A specific embodiment 1 of the present application is provided below, and the specific implementation of each step in embodiment 1 is described in detail as follows.

[0154] The specific implementation of steps S01-S02 is the same as the foregoing, and will not be described again here.

[0155] The specific implementation of step S03 is to perform a first-level classification of errors. First, the error feature data is preliminarily clustered using a K-means clustering algorithm, and the number of clusters is set to 3, representing three main error types. The objective function of K-means clustering can be expressed as: In the formula, J is the clustering objective function; E i is the feature vector of the i-th error sample; C j is the j-th cluster center; and N is the total number of error samples. Then, based on the time stability index, spectral characteristics, and environmental correlation coefficient, etc., the error values are classified into three categories: systematic error values, random error values, and environmental interference error values. The time stability index calculation formula is: In the formula, S t is the time stability index; σ 2 (t) is the variance of the error at time t; σ 2 (t+Δt) is the variance of the error at time t+Δt; and Δt is the time interval. When S t <0.05%, it is identified as a systematic error. The signal-to-noise ratio calculation formula of the spectral characteristics is: In the formula, SNR is the signal-to-noise ratio, with the unit of decibel (dB); P signal is the signal power; and P noise is the noise power. When SNR<15dB and contains high-frequency components, it is identified as a random error. The environmental correlation coefficient calculation formula is: In the formula, ρ E，P is the correlation coefficient of the error and the environmental parameter; Cov(E, P) is the covariance of the error E and the environmental parameter P; σ E is the standard deviation of the error; and σ P is the standard deviation of the environmental parameter. When |ρ E，P |>0.7, it is identified as an environmental interference error. For boundary cases that are difficult to directly classify, fuzzy logic rules are used for judgment, and the membership degree function calculation formula is: In the formula, μ j (E i ) is the membership degree of the error sample E i to the j-th error; and ||E i -C j|| is the Euclidean distance from the error sample to the cluster center; d0is the standardization parameter, which is the average distance between samples; p is the fuzzy factor, which is 2. The type corresponding to the highest probability is taken as the classification result. This classification process utilizes the essential characteristics of errors for classification, laying a foundation for subsequent more refined classification and targeted compensation.

[0156] The specific implementation of step S04 is to perform secondary classification of errors. For systematic error values, classification is performed by linear correlation test, and the Pearson correlation coefficient calculation formula is: In the formula, r is the Pearson correlation coefficient; E i is the i-th error value; is the error average value; M i is the i-th measurement value; is the measurement average value; and n is the sample quantity. When |r|>0.85 and the error changes with the measurement value in a clear linear relationship, it is classified as a linear error value; when |r|<0.85 or the error and the measurement value present a clear nonlinear characteristic, it is classified as a nonlinear error value. For random error values, classification is performed by frequency domain analysis, and the fast Fourier transform calculation formula is: In the formula, X k is the frequency domain representation; x n is the time domain error sequence; N is the sequence length; k is the frequency index; and i is the imaginary unit. The high-frequency energy proportion calculation formula is: In the formula, R HF is the high-frequency energy proportion; |X k | is the frequency domain amplitude; and N is the FFT point number. When the spectral energy is mainly concentrated in the high-frequency band (more than 1 / 4 of the sampling frequency), it is classified as a noise error value; when the spectral energy is widely distributed and there is a clear low-frequency fluctuation characteristic, it is classified as a fluctuation error value. For environmental interference error values, classification is performed by environmental sensitivity analysis, and the temperature sensitivity calculation formula is: In the formula, S T is the temperature sensitivity; is the partial derivative of the error with respect to temperature; ΔE is the error change; and ΔT is the temperature change. When S T >0.1% full scale / °C, it is classified as a temperature error value; when the error and the salinity measurement value or the conductivity exist cross-influence relationship, it is classified as a salinity mutual interference error value. This secondary classification method utilizes the mathematical characteristics and physical mechanisms of errors for more accurate classification, providing a direction for subsequent refined compensation.

[0157] The specific implementation of step S05 is to perform three-level classification of errors. For linear error values, a linear regression equation of error and measurement value is analyzed: E = β0 + β1M + ε, where E is the error value; β0 is the intercept, representing the zero-point drift; β1 is the slope, representing the range offset; M is the measurement value; and ε is the random error term. When |β0| > 0.3% full scale, it is classified as a zero-point drift error value; when |β1-1| > 0.05, it is classified as a range offset error value. For nonlinear error values, the hysteresis error calculation formula is obtained by comparing the rising and falling measurements: E hys = |E up (M) - E down (M) |, where E hys is the hysteresis error value; E up (M) is the error when the measurement value is M in the rising process; and E down (M) is the error when the measurement value is M in the falling process. When E hys > 0.2% full scale, it is classified as a hysteresis error value. The derivative comparison formula for saturation error determination is: where R sat is the derivative ratio; dE / dM bound is the derivative of error with respect to measurement value near the range boundary; and dE / dM normal is the derivative of error with respect to measurement value in the normal region. When R sat > 2, it is classified as a saturation error value. For noise error values, through time-frequency analysis, the pulse noise determination formula is: where P i is the pulse flag; E i is the i-th error value; μ is the error mean; σ is the error standard deviation; and d i is the abnormal duration (number of sampling periods). When the error contains a persistent high-frequency component with a frequency higher than 1 / 3 of the measurement frequency, it is classified as a high-frequency noise error value; when there is an abnormal point in the error with an amplitude greater than 3 times the standard deviation and a duration shorter than 3 sampling periods, it is classified as a pulse noise error value. For fluctuation error values, through autocorrelation analysis, the autocorrelation function calculation formula is: where R xx (k) is the autocorrelation function with a time delay of k; E i is the i-th error value; μ is the error mean; and N is the sequence length. The random jump determination formula is: where J i is the jump flag; E i is the i-th error value; is the average error variation amplitude. When the autocorrelation function of the error presents obvious periodicity and the period is stable, it is classified as a periodic fluctuation error value; when the error has irregular mutations and the variation amplitude is greater than 2 times the average fluctuation amplitude, it is classified as a random jump error value. For the temperature error value, through multi-point temperature measurement analysis, the temperature gradient correlation coefficient calculation formula is: wherein, is the correlation coefficient of the error and the temperature gradient; is the covariance of the error E and the temperature gradient ; σ E is the standard deviation of the error; is the standard deviation of the temperature gradient. The temperature compensation insufficient error determination formula is: wherein, R TC is the correlation coefficient of the residual error after temperature compensation and the temperature; E comp is the residual error after temperature compensation; T is the temperature value; Cov(E comp , T) is the covariance of the error after compensation and the temperature; is the standard deviation of the error after compensation; σ T is the standard deviation of the temperature. When , it is classified as a temperature gradient error value; when the residual error after temperature compensation still presents correlation with the temperature change, it is classified as a temperature compensation insufficient error value. For the salinity mutual interference error value, through ion selectivity test, when the error changes with the change of the concentration of a specific ion, it is classified as an ion interference error value; when the error presents nonlinear growth under high salinity conditions (greater than 30 psu), it is classified as a conductivity nonlinear error value. The three-level classification process uses the detailed characteristics of the error for accurate classification, laying a foundation for subsequent accurate compensation.

[0158] The specific implementation of step S06 is to call an optimal path selection function to determine the best compensation strategy. The function uses a decision fusion algorithm based on evidence theory to determine the optimal compensation path by comprehensively considering multi-source information. First, singular spectrum analysis is applied to the input small-sized thermosalinograph measurement data to construct a trajectory matrix: wherein, X is the trajectory matrix; x i is the time series data point; L is the window length; K = N-L+1, N is the sequence length. Then singular value decomposition is performed: X = U∑V T , wherein, U is the left singular vector matrix; ∑ is the singular value diagonal matrix; V T is the transpose of the right singular vector matrix. The main mode characteristics and abnormal characteristics of the data are extracted, and the similarity between the characteristic vector and each type of error characteristic is calculated: wherein, S(F, F j ) is the cosine similarity between the characteristic vector F and the error characteristic F j ; F·Fj is the dot product of two vectors; ||F|| and ||F j are the Euclidean norms of two vectors. Then, the environment parameter similarity matrix is constructed, and the Euclidean distance calculation formula is: where D ij is the Euclidean distance between the current environment parameter i and the historical environment parameter j; P ik is the kth component of the current environment parameter vector; P jk is the kth component of the historical environment parameter vector; and m is the dimension of the environment parameter. The threshold is set to 0.15. Next, the pattern matching of historical error statistics data is performed, and the dynamic time warping algorithm is used, and the distance matrix calculation formula is:

[0159] DTW(i, j) = d(i, j) + min{DTW(i-1, j), DTW(i, j-1), DTW(i-1, j-1)}, where DTW(i, j) is the minimum matching distance of the first i points and the first j points; and d(i, j) is the distance between the ith point and the jth point. The threshold is set to 0.75. At the same time, the influence factor is calculated according to the device state information, and the comprehensive influence factor calculation formula is: device F b = w b ·F s +w s ·F c +w c ·F device , where F b is the device state comprehensive influence factor; F s is the battery power attenuation coefficient, which is 2 when the power is less than 30%, and otherwise is 1; F c is the sensor aging coefficient, which is 1.5 when the usage time is more than 500 hours, and otherwise is 1; F b is the calibration time coefficient, which is 1.8 when the distance to the last calibration is more than 30 days, and otherwise is 1; w s , w c , and w i are the weights of the three coefficients, and the default is 1 / 3. Finally, considering the measurement mode identifier, different compensation strategy weights are assigned to different modes (static, dynamic, continuous monitoring, and fixed point sampling). Based on the above analysis results, the analytic hierarchy process is used to calculate the comprehensive score of each candidate compensation path: where Score j is the comprehensive score of the ith compensation path; w ijLet be the score of the i-th path on the j-th indicator; n is the number of evaluation indicators. The path with the highest score is selected as the optimal neural network compensation path. This method considers multi-dimensional factors and can adaptively select the compensation strategy most suitable for the current measurement conditions, improving the accuracy and stability of the compensation effect.

[0160] The specific implementation method of step S07 is the same as described above, and will not be repeated here.

[0161] The specific implementation of step S08 involves applying a compensation chain to perform multi-step error correction. The raw measurement data from the small temperature and salinity meter and the current environmental parameters are used as inputs and fed into the first model of the compensation chain. Each model receives the output (or raw input) of the previous model, processes it, calculates the error compensation amount for a specific type, and passes the corrected data to the next model. The system linear error compensation calculation formula is: T comp1 =T raw -C zero，T +T raw ·C span，T S comp1 =S raw -C zero，S +S raw ·C span，S In the formula, T comp1 and S comp1 These are the temperature and salinity values ​​after system linearity error compensation, respectively; T raw and S raw The original measured temperature and salinity values; C zero，T and C zero，S These are the zero-point drift error compensation amounts for temperature and salinity, respectively; C span，T and C span，T These are the range offset error compensation amounts for temperature and salinity, respectively. The system nonlinearity error compensation calculation formula is: T comp2 =T comp1 -C hys，T (T comp1 ,ΔT / Δt)-C sat，T (T comp1 ), S comp2 =S comp1 -C hys，S (S comp1 ,ΔS / Δt)-C sat，S (S comp1 In the formula, T comp2 and S comp2 These are the temperature and salinity values ​​after system nonlinear error compensation, respectively; C hys，T and C hys，S C represents the hysteresis error compensation functions for temperature and salinity, respectively, and is related to the measured values ​​and their rates of change; sat，T and Csat，S are the saturation error compensation functions of temperature and salinity, respectively, which are related to the measured values. The wavelet transform calculation formula of random noise error compensation is: where W j，k is the wavelet coefficient; x(t) is the signal; ψ is the wavelet base function; j is the scale parameter; and k is the translation parameter. The wavelet threshold denoising calculation formula is: where W is the wavelet coefficient after threshold processing;

[0162] W j，k is the original wavelet coefficient; λ j is the threshold value corresponding to the scale j, which is usually taken as where σ j is the noise standard deviation of the scale j, and N is the signal length. The GRU calculation formula of random fluctuation error compensation is: z t =σ(W z ·[h t-1 , x t ]+b z ), r t =σ(W r ·[h t-1 , x t ]+b r ), where z t is the update gate; r t is the reset gate; is the candidate hidden state; h t is the current hidden state; x t is the input sequence; W z , W r , W z are the weight matrices; b r , b comp4 , b comp4 are the bias vectors; σ is the sigmoid activation function; and tanh is the hyperbolic tangent activation function. The environmental temperature error compensation calculation formula is: where T comp4 and S comp4 are the temperature value and salinity value after environmental temperature error compensation, respectively; is the error compensation amount caused by the temperature gradient, which is a function of the temperature gradient ; C insuf (T, ΔT) is the error compensation amount caused by the insufficient temperature compensation, which is related to the temperature T and its change ΔT; is the compensation amount of the influence of temperature on salinity measurement. The environmental salinity mutual interference error compensation calculation formula is: S comp5 =S comp4 -C ion(I1, I2,..., I n )-C nonlin (EC, S comp4 ), wherein S comp5 is the salinity value after compensation of environmental salinity interference error; C ion (I1, I2,..., I n ) is the ion interference error compensation amount, which is a function of various ion concentrations I1, I2,..., I n ; C nonlin (EC, S comp4 ) is the conductivity nonlinear error compensation amount, which is related to the conductivity EC and the compensated salinity value S comp4 . The final output calculation formula is: T final = T comp4 , S final = S comp5 , wherein T final and S final are the final temperature output value and salinity output value after the complete compensation chain processing, respectively. During the entire process, an error propagation control strategy is adopted to ensure that each level of compensation is coordinated with each other, and to avoid error accumulation and overcompensation. This method greatly improves the measurement accuracy through multi-stage fine compensation, so that the small-sized CTD can still achieve measurement performance close to high-end equipment under low-power consumption conditions.

[0163] Optionally, the specific implementation of step S09 is to periodically update the error feature database to adapt to the dynamic changes of the equipment and the environment. First, an automatic data acquisition mechanism is established to store measurement data, environmental parameters and corresponding error analysis results periodically (such as every 100 measurements) during normal measurement. Then, a data screening algorithm is designed to exclude unreliable data through anomaly detection technology and retain representative error samples. The isolation forest algorithm for anomaly detection calculates the anomaly score formula as follows: wherein s(x, n) is the anomaly score of sample x; E(h(x)) is the average path length of sample x; c(n) is the average path length of the binary search tree with sample number n, usually c(n) = 2H(n-1)-(2(n-1) / n), wherein H(i) is the harmonic number. When s(x, n) > 1.5% of the full range, the sample is considered as an outlier. Then, an incremental learning strategy is implemented to integrate new data with the existing database, and a weighted average method is used to update the error feature statistical parameters. The weighted average calculation formula is: F new = (1-α)·F old +α·F recent , wherein F new is the updated error feature parameter; F old is the original error feature parameter; F recentis the error characteristic parameter of the newly collected data; a is the new data weight coefficient, and the value is 0.3. At the same time, the time decay mechanism is applied to gradually weaken the influence of the old data over time, and the time decay calculation formula is as follows: In the formula, w i is the weight of the ith data; λ is the decay coefficient, and the value is 0.05 / month; t i is the number of months from the current time. When the cumulative amount of new data exceeds 10% of the database capacity or a significant change in the error characteristics is detected (such as a change in the average error of a certain type of error exceeding 30%), the database reconstruction process is triggered, the index structure is reorganized, and the storage is optimized. Finally, the updated error characteristic database is automatically synchronized to the optimal path selection function to ensure that the path selection is always based on the latest error characteristics. The implementation adopts an incremental database maintenance strategy to realize dynamic updating and adaptive evolution of the error characteristic library, so that the compensation system can timely cope with new error patterns caused by factors such as equipment aging and environmental changes, and maintain long-term compensation effect.

[0164] In order to better understand and implement the present application, the following provides an embodiment 2 of a specific application scenario of the present application: when researchers use a small temperature salinity instrument to monitor marine environmental parameters in a certain marine scientific expedition, they find that due to the power consumption limitation of the instrument and the complex marine environment, the measurement data has obvious errors. In order to improve the data accuracy, the researchers applied the multi-level neural network tree structure error compensation method of the present application. In this example, the temperature salinity instrument is powered by a 12V lithium battery, with a rated power consumption of 0.8W, and a measurement range of temperature-5℃ to 45℃ and salinity 0 to 42psu.

[0165] Firstly, the researchers constructed a multi-level neural network tree structure, including three layers of classification nodes and one layer of leaf nodes. The root node is an error classifier, the second layer sets three nodes of systematic error, random error and environmental interference error, the third layer sets six nodes of linear error, nonlinear error, etc., and the fourth layer sets twelve leaf nodes of zero drift error, etc. Each node is a miniature neural network with a parameter amount less than 100, ensuring efficient operation under the limited computing resources of the temperature salinity instrument.

[0166] In the data collection stage, the researchers placed the small temperature salinity instrument and the high-precision standard instrument (precision: temperature ±0.002℃, salinity ±0.003psu) in different sea areas for comparison measurement, including nearshore waters, deep sea areas and special water chemical environments, a total of 8000 data points were collected, forming a complete error characteristic database. Part of the original data collected is shown in Table 1:

[0167] Table 1 Comparison data of measurement of small temperature salinity instrument and standard instrument in different environments

[0168]

[0169]

[0170] In the first-level classification, researchers applied the K-means clustering algorithm to classify errors into systematic, random, and environmental interference errors. The clustering results showed that systematic errors accounted for about 58% of the total errors, random errors accounted for about 28%, and environmental interference errors accounted for about 14%. By calculating the time stability index, spectral characteristics, and environmental correlation coefficient, the rationality of the classification results was further verified.

[0171] In the second and third-level classification processes, researchers applied various analysis methods to refine error types. For example, through Pearson correlation coefficient calculation, it was found that the systematic error in temperature measurement was mainly linear error (r=0.87), while the systematic error in salinity measurement contained obvious nonlinear components (r=0.73). Through FFT frequency domain analysis, it was found that the proportion of high-frequency noise in random error was 62%, and the proportion of fluctuation error was 38%. The temperature gradient sensitivity calculation result was 0.13% full scale / ℃, indicating that the change of environmental temperature had a significant impact on the measurement results. The proportion of each type of error is shown in Table 2:

[0172] Table 2 Proportion of each type of error in total error

[0173]

[0174] Based on the error classification results, researchers called the optimal path selection function to determine the compensation strategy. This function considered feature similarity, environmental parameter changes, historical error patterns, and device state information. For example, when measuring in the thermocline area (temperature gradient > 0.5℃ / m), the optimal compensation path prioritizes handling temperature gradient errors; while in the estuary mixing zone, the optimal path focuses on handling ion interference errors.

[0175] Researchers selected corresponding models from the pre-trained micro neural network library to construct the compensation chain. The data of three typical sea areas were processed using the compensation chain, and the compensation effect is shown in Table 3:

[0176] Table 3 Comparison of data accuracy before and after error compensation

[0177]

[0178]

[0179] The researchers conducted a long-term test of the compensation system for 3 months, during which the error feature database was updated every week. The outliers were identified by the isolation forest algorithm, and the outlier score threshold was set to 1.5% full range, and about 3.8% of the abnormal data were removed. The weighted average method (new data weight α = 0.3) and time decay mechanism (decay coefficient λ = 0.05 / month) were used to update the feature parameters, ensuring that the compensation system can adapt to equipment aging and environmental changes. The test results show that the long-term stability of the compensation system is good, and the accuracy does not decrease significantly during the 3-month test.

[0180] The traditional error compensation method of the conventional CTD instrument mainly adopts the combination of fixed formula calibration and simple digital filtering, which can only deal with linear system error and simple random noise, and is almost powerless to non-linear error and environmental interference. These methods usually need to be calibrated frequently under standard environment, and once the environmental conditions change or the equipment ages, the measurement accuracy will decrease significantly. The multi-level neural network tree structure error compensation method adopted by the present application can accurately identify and classify various error types and apply targeted compensation strategies. In addition, the present application realizes adaptive compensation by regularly updating the error feature database, greatly prolongs the calibration period of the instrument, reduces the maintenance cost, and enables the small CTD instrument to still obtain high-precision measurement results under low-energy consumption conditions.

[0181] It should be noted that the variables involved in the present application are explained in detail as shown in Tables 4, 5 and 6.

[0182] Table 4 Variable Explanation Table (First Part)

[0183]

[0184]

[0185] Table 5 Variable Explanation Table (Second Part)

[0186]

[0187] Table 6 Variable Explanation Table (Third Part)

[0188]

[0189]

Claims

1. A method for compensating for data acquisition errors in a small-scale thermo-salinity meter, characterized in that, include: Construct a multi-level neural network tree structure, with each node set as a pre-trained miniature neural network. The number of tree paths is determined based on the number of error types. Collect measurement data from a small temperature and salinity meter and compare it with standard data to calculate error values, establishing an error feature database. Error features are used for primary classification; Systematic error values, random error values, and environmental interference error values ​​are classified into two-level and three-level categories, respectively. The optimal path selection function is called, taking into account the measurement data of the small temperature and salinity meter, environmental parameter values, historical error statistics, equipment status information, and measurement mode identifier, and outputting the optimal neural network compensation path. Based on the optimal neural network compensation path, the corresponding error compensation model is selected from the pre-trained micro-neural network library, and multiple micro-neural networks are connected in series to form a compensation chain. The compensation chain is applied to perform multi-step error correction on the measurement data of the small temperature and salinity meter, and the corrected temperature and salinity values ​​are output.

2. The method according to claim 1, characterized in that, The primary classification based on error characteristics is as follows: error values ​​are divided into systematic error values, random error values, and environmental interference error values. The secondary classification of systematic error values ​​includes linear error values ​​and nonlinear error values. The secondary classification of random error values ​​includes noise error values ​​and fluctuation error values. The secondary classification of environmental interference error values ​​includes temperature error values ​​and salinity mutual interference error values.

3. The method according to claim 2, characterized in that, The three-level classification of linear error values ​​includes zero-point drift error values ​​and range offset error values; the three-level classification of nonlinear error values ​​includes hysteresis error values ​​and saturation error values; the three-level classification of noise error values ​​includes high-frequency noise error values ​​and impulse noise error values; the three-level classification of fluctuation error values ​​includes periodic fluctuation error values ​​and random jump error values; the three-level classification of temperature error values ​​includes temperature gradient error values ​​and insufficient temperature compensation error values; and the three-level classification of salinity interference error values ​​includes ion interference error values ​​and conductivity nonlinearity error values.

4. The method according to claim 3, characterized in that, The selected error compensation models include system linear error compensation model, system nonlinear error compensation model, random noise error compensation model, random fluctuation error compensation model, environmental temperature error compensation model, and environmental salinity mutual interference error compensation model.

5. The method according to claim 4, characterized in that, The specific structure of the system linear error compensation model is a three-layer feedforward neural network. The input layer has two neurons that receive the original measured values ​​of temperature and salinity. The hidden layer has four neurons that use the Sigmoid activation function. The output layer has two neurons that output the compensation amount for zero-point drift error and the compensation amount for range offset error, respectively. The total number of parameters in the system linear error compensation model is controlled within 50.

6. The method according to claim 5, characterized in that, The specific structure of the system nonlinear error compensation model is a four-layer feedforward neural network. The input layer has three neurons that receive temperature, salinity, and rate of change values. The hidden layers contain six and four neurons respectively. The output layer has two neurons that output the compensation amount for hysteresis error and saturation error respectively. The system nonlinear error compensation model adopts a sparse connection structure, and the total number of parameters is controlled within 80.

7. The method according to claim 6, characterized in that, The specific structure of the random noise error compensation model is a recurrent neural network based on wavelet transform. The input layer receives five consecutive temperature and salinity measurement data points, and includes a denoising layer and a signal reconstruction layer, for a total of three layers. The number of parameters is controlled within 60. The random noise error compensation model has the ability to filter out high-frequency noise error values ​​and identify impulse noise error values.

8. The method according to claim 7, characterized in that, The specific structure of the random fluctuation error compensation model is a gated recurrent unit network. The input layer receives 10 consecutive temperature and salinity measurement data points, the hidden layer contains 3 GRU units, and the output layer has 2 neurons that output the compensation amount of periodic fluctuation error value and the detection threshold of random jump error value, respectively. The total number of parameters of the random fluctuation error compensation model is controlled within 70.

9. The method according to claim 8, characterized in that, The specific structure of the ambient temperature error compensation model is a three-layer feedforward neural network. The input layer receives the measured temperature values ​​and their gradient information values ​​from multiple points. The hidden layer has 5 neurons. The output layer outputs the compensation amount for the temperature gradient error value and the correction amount for the temperature compensation deficiency error value. The total number of parameters in the ambient temperature error compensation model is controlled within 60.

10. The method according to claim 9, characterized in that, The specific structure of the environmental salinity mutual interference error compensation model is a four-layer feedforward neural network. The input layer receives the measured salinity value, conductivity value, and various ion concentration estimates. The two hidden layers contain 6 and 4 neurons respectively. The output layer outputs the compensation amount for ion interference error and the correction amount for conductivity nonlinearity error. The total number of parameters in the environmental salinity mutual interference error compensation model is controlled within 90.

Images