A marine multi-parameter precision measuring instrument operation fault intelligent early warning method

Abstract

This invention provides an intelligent early warning method for operational faults of a multi-parameter precision measuring instrument in the ocean, belonging to the field of marine measurement technology. The invention acquires the time series data of the multi-parameter measuring instrument and inputs it into a Gaussian process regression drift compensation model to obtain a compensated sensor observation sequence. A comprehensive early warning index is used to trigger the corresponding early warning level, and the inference depth and sampling frequency are adaptively adjusted. For missing data, a physical constraint repair method based on a diffusion model is used to update the anomaly score. Finally, the model is compressed into a lightweight student model through knowledge distillation, quantization, and pruning, and deployed on an underwater embedded processor. By utilizing the collaborative mechanism of artificial intelligence and physical equation constraints, the technical problem of achieving physically consistent perception and node-level positioning of multi-sensor fault information within a unified framework is solved.

Description

Technical Field

[0001] This invention belongs to the field of marine measurement technology, and more specifically, relates to an intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument. Background Technology

[0002] Marine multi-parameter precision measuring instruments operate long-term in the high-pressure, high-salinity, and low-temperature deep-sea environment, requiring the simultaneous acquisition of multiple time series data, including temperature, pressure, salinity, dissolved oxygen, and power supply voltage. Traditional fault early warning methods primarily rely on single-sensor threshold judgments, supplemented by polynomial temperature and pressure compensation models to correct sensor drift, with fault triggering achieved through manually set fixed alarm thresholds. However, in current deep-sea multi-parameter monitoring, due to individual differences in sensor aging patterns, fixed polynomial compensation models struggle to track non-stationary drift. Furthermore, single-threshold judgments neglect the physical coupling relationships between multiple sensors, failing to distinguish between global and local faults. Existing methods lack physical constraint repair mechanisms for missing data, leading to data quality degradation after telemetry interruptions, further impacting fault judgment accuracy. In other words, current technologies suffer from the technical challenge of achieving physically consistent perception and node-level localization of multi-sensor fault information within a unified framework. Summary of the Invention

[0003] In view of this, the present invention provides an intelligent early warning method for operational faults of marine multi-parameter precision measuring instruments, which can solve the technical problem in the prior art that multi-sensor fault information cannot achieve physical consistency perception and node-level positioning under a unified framework.

[0004] This invention is implemented as follows: This invention provides an intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument, comprising the following steps:

[0005] The time series of temperature, pressure, salinity, dissolved oxygen and power supply voltage of the multi-parameter measuring instrument are collected, input into the Gaussian process regression drift compensation model, and output the compensated sensor observation sequence of each sensor.

[0006] The compensated sensor observation sequence is input into the spectrogram wavelet transform module. The edge weights are determined by the Granger causality test, the graph Laplacian matrix is ​​constructed, the graph-time joint time-frequency representation is output, and the fault source node is located in reverse.

[0007] The residual sequences of the temperature-salinity-density equation and the dissolved oxygen saturation equation are calculated in real time and input together with the compensated sensor observation sequence into the dual-stream contrast self-supervised converter model to output the physical consistency perception anomaly score.

[0008] The comprehensive early warning index is calculated by taking the physical consistency perception anomaly score, the proportion of anomaly energy in the graph-time joint time-frequency representation, and the mean of the physical consistency perception anomaly score in the most recent sampling point as inputs. The corresponding early warning level is triggered according to the interval to which the comprehensive early warning index belongs, and the inference calculation depth and sampling frequency are adjusted accordingly.

[0009] For telemetry time series with missing segments, a time series missing data repair method based on diffusion model is used in combination with physical constraints to repair the data. The repaired telemetry time series is then input into a dual-stream comparison self-supervised transformer model to update the physical consistency perception anomaly score.

[0010] The dual-stream contrastive self-supervised transformer model is compressed into a lightweight student model through knowledge distillation, quantization, and structured pruning. It is then deployed on an underwater embedded processor and the CMSIS-NN library is used to accelerate instruction-level inference on the ARM architecture, enabling real-time early warning at the device end.

[0011] The Gaussian process regression drift compensation model uses temperature, pressure, and aging time as three-dimensional inputs and sensor baseline drift as the output target. It constructs a Gaussian process prior using a squared exponential kernel function, iteratively updates the kernel function hyperparameters using historical online calibration data, and maintains a set of model parameters independently for each sensor.

[0012] The baseline drift of the sensor is the difference between the actual output value of each sensor under the current temperature, pressure, and aging time conditions and the factory calibration value, which is obtained by fitting historical online calibration data.

[0013] In this process, the Granger causality test establishes univariate autoregressive models and bivariate autoregressive models for the sensor pairs respectively, compares the degree of decrease in prediction error, establishes directed edges for sensor pairs whose decrease exceeds the significance threshold, and uses the decrease in prediction error as the edge weight of the graph Laplacian matrix.

[0014] The graph Laplacian matrix is ​​obtained by subtracting the degree matrix and adjacency matrix of the weighted directed graph. After eigenvalue decomposition of the graph Laplacian matrix, a spectral wavelet basis is constructed. The compensated sensor observation sequence is decomposed along the graph domain and time domain in two dimensions to obtain the graph-time joint time-frequency representation.

[0015] The dual-stream contrastive self-supervised transformer model includes a physical stream encoder, a semantic stream encoder, a dynamic routing module, a contrastive learning head, and an anomaly detection head. The physical stream encoder receives the compensated sensor observation sequence, and the semantic stream encoder receives the concatenated vector of the residual sequence of the temperature-salinity-density equation and the residual sequence of the dissolved oxygen saturation equation. Both encoders use rotational position encoding.

[0016] The dynamic routing module calculates the cosine similarity between the physical flow feature vector and the semantic flow feature vector. When the cosine similarity is lower than the dynamic routing trigger threshold, the cross-flow attention module is activated to perform bidirectional interaction between the physical flow feature vector and the semantic flow feature vector.

[0017] The contrastive learning head uses the physical flow feature vector and semantic flow feature vector at the same time to form positive sample pairs, and uses feature vectors at other times to form negative samples. The InfoNCE loss function is used to guide the encoder to learn a compact representation of physical consistency awareness.

[0018] The comprehensive early warning index is calculated by weighted summation of the physical consistency perception anomaly score, the proportion of anomaly energy in the graph-time joint time-frequency representation, and the mean of the physical consistency perception anomaly score within the most recent sampling point. The weights are determined by experiments on the historical validation set that minimize the sum of the false negative rate and the false positive rate.

[0019] Specifically, when the comprehensive early warning index is below the first early warning threshold, normal monitoring is maintained; when the comprehensive early warning index is between the first and second early warning thresholds, a level one early warning is triggered, activating the cross-flow attention module and increasing the sampling frequency; when the comprehensive early warning index is between the second and third early warning thresholds, a level two early warning is triggered, sending a fault risk warning to the shore-based system and initiating the fault source node location process; when the comprehensive early warning index is not lower than the third early warning threshold, a level three early warning is triggered, sending an emergency alarm to the shore-based system and suspending the acquisition of non-critical parameter channels.

[0020] The time series missing data repair method based on the diffusion model takes the Gaussian noise addition process of the complete time series as the forward process and the denoising score matching as the training target. In the reverse process, the consistency residual of the temperature-salt-density equation is used as a physical constraint to guide the denoising direction.

[0021] Specifically, for missing segments with consecutive missing lengths exceeding the missing length threshold, a time series missing data repair method based on a diffusion model is used for repair; for missing segments with consecutive missing lengths not exceeding the missing length threshold, cubic spline interpolation is used.

[0022] The knowledge distillation process uses a dual-stream contrastive self-supervised transformer model as the teacher model, a lightweight convolutional network with no more than the parameter compression ratio threshold of the teacher model as the lightweight student model, and the mean square error between the intermediate layer feature maps of the teacher model and the corresponding layer feature maps of the lightweight student model as the distillation loss. The total training loss is obtained by weighted summation of the anomaly detection classification loss of the lightweight student model.

[0023] The quantization process quantizes the floating-point weights and activation values ​​of the lightweight student model to an 8-bit integer representation, and the quantization calibration dataset consists of normally functioning samples. The structured pruning process uses the L1 norm of each convolutional kernel output channel as an importance score, and prunes channels with scores lower than the pruning threshold.

[0024] The search range for the dynamic routing trigger threshold is 0.5–0.95, with a step size of 0.05; the weight range for the comprehensive early warning index is 0.1–0.8, with a step size of 0.05; the first early warning threshold is 0.3, the second early warning threshold is 0.6, and the third early warning threshold is 0.85; the missing length threshold is 10 sampling points; the compression ratio threshold is 2% of the teacher model parameters; the number of most recent sampling points is 60; and the pruning threshold is gradually determined with the constraint that the accuracy of the physical consistency perception anomaly score decreases by no more than 1%.

[0025] This invention achieves individualized adaptive baseline correction by constructing a Gaussian process regression drift compensation model, eliminating the interference of sensor aging drift on fault diagnosis. It utilizes spectral wavelet transform to fuse multi-sensor spatial topology and temporal dynamics into a graph-time joint time-frequency representation, using Granger causality tests to determine edge weights, enabling fault propagation path tracing and fault localization at the node level, avoiding misjudgments of fault sources caused by single-sensor threshold judgments. Furthermore, it employs a dual-stream contrastive self-supervised transformer model to synchronously perceive the consistency between the original dynamic characteristics and physical equations of the sensors, adaptively adjusts the inference depth and sampling frequency using a comprehensive early warning index, and ensures the physical rationality of missing data segments using a diffusion model-based repair method. In summary, this invention solves the technical problem mentioned in the background art where multi-sensor fault information cannot achieve physically consistent perception and node-level localization within a unified framework. Attached Figure Description

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

[0027] Figure 2 This is a graph showing the changes in the comprehensive early warning index over time and the triggering sequence of each early warning level.

[0028] Figure 3 This is a comparison chart of the pressure sensor output values ​​before and after Gaussian process regression drift compensation. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0030] like Figure 1The diagram shown is a flowchart of an intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument provided by the present invention. This method includes the following steps:

[0031] S01. Collect the time series of temperature, pressure, salinity, dissolved oxygen and power supply voltage of the multi-parameter measuring instrument, input the Gaussian process regression drift compensation model, and output the compensated sensor observation sequence of each sensor.

[0032] S02. Input the compensated sensor observation sequence into the spectrogram wavelet transform module, determine the edge weights using the Granger causality test, construct the graph Laplace matrix, output the graph-time joint time-frequency representation, and inversely locate the fault source node.

[0033] S03. Real-time calculation of the residual sequence of the temperature-salinity-density equation and the residual sequence of the dissolved oxygen saturation equation, together with the compensated sensor observation sequence, is input into the dual-stream comparison self-supervised converter model, and the physical consistency perception anomaly score is output.

[0034] S04. Using the physical consistency perception anomaly score, the proportion of anomaly energy in the graph-time joint time-frequency representation, and the average physical consistency perception anomaly score in the last 60 sampling points as inputs, calculate the comprehensive early warning index, trigger the corresponding early warning level according to the interval to which the comprehensive early warning index belongs, and adjust the inference calculation depth and sampling frequency.

[0035] S05. For telemetry time series containing missing segments, a time series missing data repair method based on diffusion model is used in combination with physical constraints to repair the missing data. The repaired telemetry time series is then input into a dual-stream comparison self-supervised transformer model to update the physical consistency perception anomaly score.

[0036] S06. The dual-stream contrastive self-supervised transformer model is quantized and structured pruning and compressed into a lightweight student model after knowledge distillation and INT8 training. It is then deployed on an underwater embedded processor and the CMSIS-NN library is used to accelerate instruction-level inference on the ARM architecture, enabling real-time early warning at the device end.

[0037] The Gaussian process regression drift compensation model uses temperature, pressure, and aging time as three-dimensional inputs and sensor baseline drift as the output target. It employs a squared exponential kernel function to construct a Gaussian process prior, iteratively updates the kernel function hyperparameters using historical online calibration data, and maintains an independent set of model parameters for each sensor to achieve individualized adaptive baseline correction. The sensor baseline drift is the difference between the actual output value of each sensor under current temperature, pressure, and aging time conditions and its factory calibration value, obtained by fitting historical online calibration data. This Gaussian process regression drift compensation model does not require a pre-defined uniform temperature and pressure compensation polynomial. Through a Bayesian update mechanism, it automatically adjusts the compensation law according to changes in the sensor's aging state, ensuring that the compensated sensor observation sequence has a reliable benchmark reference value under deep-sea conditions.

[0038] The Granger causality test establishes univariate and bivariate autoregressive models for each sensor pair, comparing the degree of decrease in prediction error. Sensor pairs with a decrease exceeding a significance threshold are assigned directed edges, with the magnitude of the prediction error decrease serving as the edge weights of the graph Laplace matrix. The significance threshold is determined through multiple segmented statistical tests on historical normal operation data, taking the critical value corresponding to the minimum confidence level that keeps the false alarm rate below 2%.

[0039] The graph Laplacian matrix is ​​obtained by subtracting the degree matrix and adjacency matrix of the weighted directed graph. After eigenvalue decomposition of the graph Laplacian matrix, a spectral wavelet basis is constructed. The compensated sensor observation sequence is then subjected to two-dimensional wavelet decomposition along the graph and time domains to obtain a graph-time joint time-frequency representation. Low-frequency spectral components correspond to global system anomalies, while high-frequency spectral components correspond to local single-node anomalies. The fault energy source nodes are solved inversely using the thermonuclear equations on the graph, with a computational complexity of O(log n). This method is applicable to sensor networks with fewer than 100 nodes. The spectral wavelet transform integrates the spatial topology and temporal dynamics of a multi-sensor system into a unified two-dimensional analysis framework. This allows the propagation path of faults in the sensor network to be traced, and distinguishes between global and local faults. As a result, the fault location accuracy is improved to the node level, avoiding misjudgments of fault sources caused by relying solely on single-sensor threshold judgments.

[0040] The specific structure of the dual-stream contrastive self-supervised transformer model is as follows: the model comprises five parts: a physical flow encoder, a semantic flow encoder, a dynamic routing module, a contrastive learning head, and an anomaly detection head. The physical flow encoder receives the compensated sensor observation sequence and consists of a 6-layer Transformer encoder, each layer containing an 8-head self-attention mechanism. Position encoding uses rotational position encoding, and the hidden layer dimension is 256. The semantic flow encoder receives the concatenated vector of the residual sequences from the temperature-salinity-density equation and the dissolved oxygen saturation equation. Its structure is symmetrical to the physical flow encoder, sharing the rotational position encoding configuration, and the hidden layer dimension is also 256. Each encoder outputs a feature vector of length 256. The dynamic routing module calculates the cosine similarity between the physical flow feature vector and the semantic flow feature vector. When the cosine similarity is lower than the dynamic routing trigger threshold, the cross-flow attention module is automatically activated to perform deep bidirectional interaction between the physical flow feature vector and the semantic flow feature vector. The contrastive learning head uses physical flow feature vectors and semantic flow feature vectors from the same time step to form positive sample pairs, and randomly sampled feature vectors from other time steps to form negative samples. The InfoNCE loss function guides the encoder to learn a compact representation of physical consistency awareness. The anomaly detection head consists of a two-layer fully connected network with an output dimension of 1. The output value is the physical consistency awareness anomaly score, ranging from 0 to 1, with higher values ​​indicating a higher degree of anomaly. During the pre-training phase, the physical flow encoder and semantic flow encoder are frozen, and only the anomaly detection head is fine-tuned.

[0041] The steps for establishing the training dataset for the dual-stream contrastive self-supervised transformer model specifically include: collecting historical online operation data, and constructing an unlabeled pre-training set by removing known fault periods from the normal operation data, with a time span of no less than 180 days; slicing the data into sliding window segments of 32 to 128 sampling points according to time window lengths, and simultaneously extracting the compensated sensor observation sequence and the residual sequences of the temperature-salinity-density equation and dissolved oxygen saturation equation at the corresponding time point for each slice; manually labeling the fault period data confirmed by manual annotation to form a labeled fine-tuning set; using a conditional generative adversarial network to synthesize high-fidelity fault samples, and expanding the fault samples in the labeled fine-tuning set to between 10% and 100% of the number of normal samples, with a total sample size of no less than 5000.

[0042] The specific steps for training the dual-stream contrastive self-supervised transformer model include: In the pre-training phase, the physical stream encoder and semantic stream encoder are trained on an unlabeled pre-training set using the InfoNCE loss function, with a batch size of 128 and an initial learning rate of [value missing]. The cosine annealing strategy is used to decay the temperature to... The training run consisted of 200 epochs. After pre-training, the physical flow encoder and semantic flow encoder were frozen. The anomaly detection head was then fine-tuned using the expanded labeled fine-tuning set, employing the FocalLoss weighted loss function, a batch size of 64, and a learning rate of [missing information]. The training rounds are 50; the dynamic routing trigger threshold is determined by performing a grid search on the validation set with the goal of maximizing fault recall, with a search range of 0.5 to 0.95 and a step size of 0.05.

[0043] The dual-stream contrastive self-supervised transformer model's physical and semantic stream architecture enables the model to simultaneously perceive the original dynamic characteristics of the sensor and the consistency of the physical equations. The two information streams are forcibly mapped to the same representation space in the contrastive learning head, allowing the physical stream encoder and semantic stream encoder to learn discriminative features that distinguish normal and abnormal states without requiring a large number of labeled fault samples. The dynamic routing module enables the model to automatically deepen the feature interaction depth between the two streams when the physical consistency perception anomaly score is high, while maintaining a shallow computational path during normal operation, achieving an adaptive balance between computational resource consumption and anomaly sensitivity. Rotational position encoding gives the self-attention mechanism a stronger ability to model the periodic dependence of ocean signals.

[0044] The comprehensive early warning index is calculated by a weighted sum of the physical consistency perception anomaly score, the proportion of anomaly energy in the graph-time joint time-frequency representation, and the average physical consistency perception anomaly score over the most recent 60 sampling points. The weights are determined experimentally on a historical validation set by minimizing the sum of the false negative and false positive rates. The weight values ​​range from 0.1 to 0.8, with a step size of 0.05 used in a grid search. Let the comprehensive early warning index be... ,when At the same time, maintain normal monitoring status, and keep the inference calculation depth at a shallow level; when When this occurs, a Level 1 warning is triggered, the cross-flow attention module is activated, and the sampling frequency is increased to twice the original sampling frequency; when When this occurs, a Level 2 early warning is triggered, a fault risk alert is sent to the shore-based system, and the fault source node location process is initiated; when At this time, a Level 3 early warning is triggered, sending an emergency alarm to the shore-based system, suspending data acquisition from non-critical parameter channels, and concentrating resources on continuous inference of critical parameter channels. The critical parameter channels refer to the temperature sensor output channel, pressure sensor output channel, and power supply voltage acquisition channel; the non-critical parameter channels refer to the salinity sensor output channel and dissolved oxygen sensor output channel.

[0045] The diffusion-based time-series missing data repair method uses the Gaussian noise addition process of the complete time series as the forward process and denoising score matching as the training objective. In the reverse process, the consistency residual of the temperature-salinity-density equation is used as a physical constraint to guide the denoising direction, ensuring the repair result satisfies the physical equation constraints. For missing segments with a continuous missing length exceeding 10 sampling points, the diffusion-based time-series missing data repair method is used; for missing segments with a continuous missing length not exceeding 10 sampling points, cubic spline interpolation is used. The physical constraints serve as input conditions to guide the generation process, ensuring that the repaired telemetry time series maintains internal consistency at the physical equation level. This avoids purely statistical generation methods producing physically unreasonable repair values ​​in deep-sea multi-parameter scenarios, thereby guaranteeing the data quality of the input dual-stream contrastive self-supervised transformer model.

[0046] The knowledge distillation process uses a dual-stream contrastive self-supervised transformer model as the teacher model and a lightweight convolutional network with no more than 2% of the teacher model's parameters as the lightweight student model. The mean squared error between the intermediate feature maps of the teacher model and the corresponding feature maps of the lightweight student model is used as the distillation loss, which is weighted and summed with the anomaly detection classification loss of the lightweight student model to obtain the total training loss. The INT8 post-training quantization quantizes the floating-point weights and activation values ​​of the lightweight student model to an 8-bit integer representation. The quantization calibration dataset consists of no fewer than 500 normally functioning samples. The structured pruning uses the L1 norm of each convolutional kernel output channel as an importance score, pruning channels with scores below a pruning threshold. This pruning threshold is determined by gradually increasing the pruning ratio on the validation set, with the constraint that the accuracy of the physically consistent perceived anomaly score does not decrease by more than 1%, until the model compression ratio is no less than 50 times.

[0047] The rotational position encoding is a positional encoding method that encodes positional information into the query vector and key vector of the attention mechanism in the form of a rotation matrix. This makes the inner product result of the self-attention mechanism depend only on the relative distance between the two positions, thereby enhancing the model's ability to model periodic time series.

[0048] The InfoNCE loss function is a contrastive learning loss function that optimizes by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. It guides the encoder to learn discriminative representations by maximizing the lower bound of mutual information.

[0049] The Focal Loss weighted loss function is a classification loss function that assigns higher loss weights to hard-to-classify samples. By using a modulation factor, it reduces the contribution of easily classified normal samples to the total loss, enabling the model to focus on optimizing the ability to identify minority class fault samples under class imbalance conditions.

[0050] The conditional generative adversarial network is a generative adversarial network that takes category labels as conditional inputs. The generator synthesizes high-fidelity samples of a specified category under the constraint of the category labels, which is used to expand the number of fault samples in the labeled fine-tuning set.

[0051] The CMSIS-NN library is a neural network inference library provided by ARM that is optimized for Cortex-M series processors. It uses SIMD instructions to accelerate convolution and matrix multiplication operations in parallel, reducing the inference latency of lightweight student models on underwater embedded processors.

[0052] Optionally, the present invention also provides a computer-based approach to form an intelligent early warning system for operational failures of a marine multi-parameter precision measuring instrument. The computer is equipped with a readable storage medium that stores program instructions, which can execute the above-described method when the computer is run.

[0053] The specific implementation of step S01 is as follows: Five time series data points from the multi-parameter measuring instrument—temperature, pressure, salinity, dissolved oxygen, and power supply voltage—are collected. These data are used as inputs to construct a Gaussian process regression drift compensation model, using temperature, pressure, and aging time as three-dimensional features. The Gaussian process regression employs a squared exponential kernel function to establish a covariance prior. The kernel function hyperparameters are iteratively updated on historical online calibration data using the marginal log-likelihood maximization method. A set of kernel function hyperparameters is maintained independently for each sensor. A Bayesian update mechanism is used to automatically adjust the compensation law according to changes in the sensor's aging state, thereby achieving individualized adaptive baseline correction. The sensor baseline drift is defined as the difference between the actual output value under the current temperature, pressure, and aging time conditions and the factory calibration value, obtained by fitting historical online calibration data. The drift output of the model is subtracted from the original sensor observations to obtain the compensated sensor observation sequence, which serves as the input for subsequent modules. The purpose of this step is to eliminate baseline drift caused by sensor aging and temperature / pressure changes under deep-sea conditions, providing a reliable reference benchmark for subsequent fault detection.

[0054] The specific implementation of step S02 is as follows: The compensated sensor observation sequence is input into the spectral wavelet transform module. First, univariate autoregressive models and bivariate autoregressive models are established between sensor pairs respectively. Granger causality tests are performed based on the magnitude of the decrease in prediction error. Directed edges are established for sensor pairs whose decrease exceeds a significance threshold. The significance threshold is obtained through multiple segmented statistical tests on historical normal operation data, and the critical value corresponding to the minimum confidence level that makes the false alarm rate less than 2% is taken. Using the magnitude of the decrease in prediction error as the edge weight, a weighted directed adjacency matrix is ​​constructed. The graph Laplace matrix is ​​obtained by subtracting the degree matrix from the adjacency matrix. Eigenvalue decomposition is performed on the graph Laplace matrix to construct a spectral wavelet basis. Two-dimensional wavelet decomposition is performed on the compensated sensor observation sequence along the graph domain and time domain to obtain the graph-time joint time-frequency representation. Low-frequency spectral components correspond to global system anomalies, and high-frequency spectral components correspond to local single-node anomalies. The fault energy source node is solved inversely using the thermonuclear equation on the graph, with a computational complexity of O(log n). This method is applicable to sensor networks with fewer than 100 nodes. The purpose of this step is to fuse multi-sensor spatial topology and temporal dynamic information into a unified analysis framework to achieve fault propagation path tracing and node-level fault location.

[0055] The specific implementation of step S03 is as follows: The residual sequences of the temperature, salinity, and density equations and the dissolved oxygen saturation equation are calculated in real time. The residual sequence of the temperature, salinity, and density equation is obtained by substituting the current observed values ​​of temperature, salinity, and pressure into the internationally accepted equation of state and then subtracting the theoretical density value. The residual sequence of the dissolved oxygen saturation equation is obtained by subtracting the observed dissolved oxygen value from the theoretical value of saturated dissolved oxygen under the corresponding temperature and salinity conditions. These two residual sequences are concatenated and input into the semantic flow encoder. The compensated sensor observation sequence is input into the physical flow encoder. Each encoder outputs a feature vector of length 256. The dynamic routing module calculates the cosine similarity between the two feature vectors. When the cosine similarity is lower than the dynamic routing trigger threshold, the cross-flow attention module is activated for bidirectional interaction. The contrast learning head uses the InfoNCE loss function to guide the encoder to learn a compact representation of physical consistency perception. The anomaly detection head outputs a physical consistency perception anomaly score from a two-layer fully connected network. The purpose of this step is to simultaneously perceive the consistency between the original dynamic characteristics of the sensor and the physical equations, and to learn discriminative features under conditions of a small number of labeled samples.

[0056] The specific implementation of step S04 is as follows: Using the physical consistency perception anomaly score, the proportion of anomaly energy in the graph-time joint time-frequency representation, and the average physical consistency perception anomaly score within the most recent 60 sampling points as inputs, a comprehensive early warning index is calculated using a weighted summation method. The weights are determined by a grid search on the historical validation set with the objective of minimizing the sum of the false negative and false positive rates. The search range is 0.1–0.8, with a step size of 0.05. When the comprehensive early warning index is below 0.3, normal monitoring is maintained, and the inference calculation depth remains at a shallow path. When the comprehensive early warning index is between 0.3 and 0.6, a Level 1 early warning is triggered, activating the cross-current attention module and increasing the sampling frequency to twice the original sampling frequency. When the comprehensive early warning index is between 0.6 and 0.85, a Level 2 early warning is triggered, sending a fault risk warning to the shore-based system and initiating the fault source node location process. When the comprehensive early warning index is not lower than 0.85, a Level 3 early warning is triggered, sending an emergency alarm to the shore-based system, suspending data acquisition from the salinity sensor output channel and dissolved oxygen sensor output channel, and concentrating resources on inference from the temperature sensor output channel, pressure sensor output channel, and power supply voltage acquisition channel. The purpose of this step is to achieve an adaptive balance between computational resource consumption and anomaly sensitivity.

[0057] The specific implementation of step S05 is as follows: For telemetry time series containing missing segments, a repair method is selected according to the length of the missing segment. For missing segments with a continuous missing length of no more than 10 sampling points, cubic spline interpolation is used, and a smooth curve is fitted using the observations before and after the missing segment for interpolation. For missing segments with a continuous missing length of more than 10 sampling points, a time series missing data repair method based on a diffusion model is used. The Gaussian noise addition process of the complete time series is used as the forward process, and the denoising score matching is used as the training objective. In the reverse denoising process, the consistency residual of the temperature-salinity-density equation is used as a physical constraint to guide the generation direction, so that the repair result satisfies the physical equation constraint. After repair, the telemetry time series is input into a dual-stream contrastive self-supervised transformer model to update the physical consistency perception anomaly score. The purpose of this step is to ensure the data quality after telemetry interruption and avoid physically unreasonable repair values ​​generated by purely statistical methods.

[0058] The specific implementation of step S06 is as follows: A dual-stream contrastive self-supervised transformer model is used as the teacher model, and a lightweight convolutional network with no more than 2% of the teacher model's parameters is used as the lightweight student model. The mean squared error between the intermediate layer feature maps of the teacher model and the corresponding layer feature maps of the lightweight student model is used as the distillation loss, which is weighted and summed with the anomaly detection classification loss of the lightweight student model to obtain the total training loss, thus completing knowledge distillation. Subsequently, the floating-point weights and activation values ​​of the lightweight student model are trained using INT8 and then quantized. The quantization calibration dataset consists of no fewer than 500 normally functioning samples. Structured pruning is performed using the L1 norm of each convolutional kernel's output channel as an importance score. The pruning threshold is constrained by a 1% decrease in the accuracy of the physical consistency perception anomaly score, and the pruning ratio is gradually increased until the compression ratio is no less than 50 times. The compressed lightweight student model is deployed on an underwater embedded processor. The CMSIS-NN library is used to accelerate convolution and matrix multiplication operations in parallel using SIMD instructions for Cortex-M series processors, enabling real-time warnings on the device. The purpose of this step is to achieve embedded deployment of the model with limited computing resources.

[0059] It should be noted that the key technologies of this invention include: Gaussian process regression drift compensation technology, which tracks the individualized aging patterns of sensors through a Bayesian update mechanism, overcoming the limitation of fixed polynomial compensation models in adapting to non-stationary drift, thus providing a reliable benchmark for the compensated data; spectral wavelet transform technology, which encodes the sensor network topology into a graph Laplacian matrix, and integrates time-domain signals and graph-domain information into the same two-dimensional analysis framework through eigenvalue decomposition, enabling fault propagation paths to be tracked and node-level localization to be achieved, fundamentally changing the defect that single-sensor threshold judgment cannot distinguish fault sources; and a dual-stream contrastive self-supervised transformer model, which forces physical flow feature vectors and semantic flow feature vectors to share a representation space through the InfoNCE loss function, learning physical consistency perception and discrimination features without requiring a large number of labeled samples, and a dynamic routing module further realizing adaptive adjustment of computational depth. These three key technologies work synergistically to form a closed loop in the system at the three levels of drift compensation, topology perception, and physical consistency judgment, mutually reinforcing each other and jointly ensuring the accuracy and real-time performance of fault warnings.

[0060] It should be noted that this invention also solves the following technical problem: In deep-sea telemetry scenarios, due to the limited bandwidth of underwater acoustic communication, telemetry time series frequently exhibit continuous missing segments. Data repaired using purely statistical methods is often inconsistent at the physical equation level, leading subsequent anomaly detection modules to misjudge repair errors as faults. This invention employs a time series missing data repair method based on a diffusion model, using the consistency residual of the temperature-salinity-density equation as a physical constraint to guide the inverse denoising direction. This ensures that the repaired data satisfies multi-parameter physical coupling relationships, thereby guaranteeing that the data input to the dual-stream contrast-based self-supervised transformer model is physically consistent. This avoids repair errors contaminating anomaly detection results and solves the technical problem of misjudging faults due to physical inconsistencies in repaired telemetry missing segment data.

[0061] Specifically, the principle of this invention is as follows: The reason why this invention can solve the above-mentioned technical problems lies in the following logical chain. Gaussian process regression uses temperature, pressure, and aging time as three-dimensional inputs, and automatically tracks individualized drift patterns through a Bayesian update mechanism, ensuring that the compensated sensor observation sequence has a reliable benchmark. Spectral wavelet transform constructs a spectral wavelet basis based on the eigenvalue decomposition of the graph Laplacian matrix, integrating the sensor network topology and temporal dynamic information into the same two-dimensional analysis framework. Low-frequency components correspond to global anomalies, and high-frequency components correspond to local anomalies. Inverse solution of the thermal kernel equation enables precise location of fault energy source nodes. The dual-flow contrastive self-supervised transformer model uses the InfoNCE loss function to force the physical flow feature vector and semantic flow feature vector to map to the same representation space. The dynamic routing module automatically deepens the cross-flow interaction depth when the anomaly level increases, enabling the physical consistency perception anomaly score to comprehensively reflect both sensor dynamic characteristic deviation and physical equation violation information. The diffusion model guides inverse denoising with physical constraints, ensuring that the repaired data satisfies the consistency of the temperature, salinity, and density equations, thereby guaranteeing the quality of data across the entire link. The above-mentioned steps are logically progressive, working together to support the realization of physical consistency perception and node-level positioning goals.

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

[0063] The specific implementation method of step S01 is as follows.

[0064] Gaussian process regression drift compensation model with temperature ,pressure Aging time For three-dimensional input, the sensor baseline drift is used. The target output is the sensor baseline drift, defined as the current actual output value. Compared with the factory calibration value The difference, that is The prior data was obtained by fitting historical online calibration data. The Gaussian process prior uses a squared exponential kernel function, the formula of which is as follows:

[0065] ;

[0066] In the formula, and The first , The input vector of each calibration sample, , The corresponding temperature is expressed in °C. , The corresponding pressure is expressed in Pa. , The corresponding aging time is expressed in hours (h). The signal variance is expressed in units of 1 and 2. The squares are consistent. This is a hyperparameter for the length scale in the temperature direction, with units of °C. This is a length-scale hyperparameter in the pressure direction, with units of Pa. The length-scale hyperparameter in the aging time direction is expressed in hours (h). The terms within the kernel function exponent are dimensionless after normalization to the corresponding length scale. For noise variance, units are... Consistent, Let Kronecker function be used when hour ,otherwise Dimensionless, the units on both sides of the equation are the same as... The squares are consistent. The above hyperparameters are iteratively updated by maximizing the logarithmic marginal likelihood function of historical calibration data, and a set of parameters is maintained independently for each sensor. Compensated sensor observation sequence Calculated by the following formula:

[0067] ;

[0068] In the formula, This represents the posterior mean predicted value of the drift amount under the current operating conditions by the Gaussian process regression model. , , The units are consistent with the units of the physical quantities output by the sensor, and the dimensions on both sides of the equation are unified.

[0069] The specific implementation method of step S02 is as follows.

[0070] Granger causality test for sensor pairs We established univariate autoregressive and bivariate autoregressive models respectively, and compared the degree of reduction in prediction error. The univariate autoregressive model is as follows:

[0071] ;

[0072] The bivariate autoregressive model is as follows:

[0073] ;

[0074] In the formula, , Sensors , At any moment The compensated observations are in units consistent with the sensor's physical quantity output units. The lag order is... Sensors in a univariate model The The autoregressive coefficient is dimensionless. , Sensors in a bivariate model Self and sensors For sensors The The regression coefficient is dimensionless. , For the corresponding residual term, the units are... Consistent; dimensions are unified on both sides of the equation. The magnitude of the decrease in prediction error. Defined as:

[0075] ;

[0076] In the formula, For variance operators, the numerator and denominator have the same units. It is a dimensionless quantity, with a value range from 0 to 1. Exceeding the significance threshold At that time, establish by point to The directed edges, the edge weights are taken as follows . The critical value corresponding to the minimum confidence level that makes the false alarm rate below 2% was determined through multiple segmented statistical tests on historical normal operation data. Adjacency matrix elements (When there exists by) point to (when there is a directed edge) or (When there are no directed edges), degree matrix It is a diagonal matrix, and its diagonal elements Graph Laplace matrix degree matrix Adjacency Matrix Subtraction yields:

[0077] ;

[0078] , , All Dimensionless real-valued matrix This represents the number of sensor nodes. Perform eigenvalue decomposition:

[0079] ;

[0080] In the formula, for The eigenvector matrix has column vectors as follows: orthogonal eigenvectors, for A diagonal matrix, with diagonal elements as The corresponding eigenvalues, for The transpose of . Construct a spectral wavelet basis and perform two-dimensional wavelet decomposition on the compensated sensor observation sequence along the graph and time domains to obtain the graph-time joint time-frequency representation. Low-frequency spectral components correspond to global system anomalies, while high-frequency spectral components correspond to local single-node anomalies. The fault energy source nodes are solved by inversely applying the on-graph thermonuclear equations, with a computational complexity of O(n log n). It is suitable for sensor networks with fewer than 100 nodes.

[0081] The specific implementation method of step S03 is as follows.

[0082] residual sequence of temperature-salinity-density equation Residual sequence of dissolved oxygen saturation equation After real-time computation, the data is concatenated with the compensated sensor observation sequence and input into a dual-stream contrastive self-supervised transformer model. The dual-stream contrastive self-supervised transformer model comprises five parts: a physical stream encoder, a semantic stream encoder, a dynamic routing module, a contrastive learning head, and an anomaly detection head. The physical stream encoder consists of a 6-layer transformer encoder, each layer containing an 8-head self-attention mechanism. The hidden layer dimension is 256, and the position encoding uses rotational position encoding, making the self-attention inner product dependent only on relative positional distance, enhancing the modeling capability for periodic time series. The semantic stream encoder structure is symmetrical to the physical stream encoder and receives... and The concatenated vectors have the same hidden layer dimension of 256. The two encoders output physical flow feature vectors respectively. With semantic flow feature vectors All are 256-dimensional real-valued vectors, dimensionless. The dynamic routing module calculates... and Cosine similarity:

[0083] ;

[0084] In the formula, Let be the dot product of two vectors. and These are the Euclidean norms of the corresponding vectors. It is a dimensionless quantity, and its range of values ​​is 1. Up to 1, both sides of the equation are dimensionless, and the dimensions are unified. When Below the dynamic routing trigger threshold At that time, the cross-flow attention module is activated to... and To conduct deep two-way interaction, The search criteria were determined by performing a grid search on the validation set with the objective of maximizing fault recall, ranging from 0.5 to 0.95 with a step size of 0.05. The contrastive learning head employed an information-noise contrastive estimation loss function.

[0085] ;

[0086] In the formula, The cosine similarity function is used. For the current moment The physical flow feature vector, For the current moment The semantic flow feature vectors, together with the two, constitute a positive sample pair. For the first The semantic flow feature vector of each negative sample. The number of negative samples. This is the temperature hyperparameter, with a default value of 0.07, dimensionless. The loss value is a dimensionless scalar value; both sides of the equation are dimensionless, and the dimensions are unified. The anomaly detection head in the fine-tuning stage uses a focal loss weighted loss function:

[0087] ;

[0088] In the formula, The probability of the model predicting the target class is dimensionless. This is the category balance coefficient, dimensionless, with an empirical value of 0.25. The modulation factor exponent is dimensionless and has an empirical value of 2. This is a modulation factor used to reduce the proportion of easily classified normal samples contributing to the total loss. The loss value is dimensionless; both sides of the equation are dimensionless, and the dimensions are unified. During the training dataset setup phase, a conditional generative adversarial network is used to synthesize high-fidelity fault samples. In category label Synthesize samples of a specified class under constraints ,in It is a random noise vector. For fault category labels, the discriminator The authenticity of a sample is determined by the joint input of the sample and the category label (here, the sample is used as the standard input). The degree matrix differs from that in step S02. The number of faulty samples is expanded to between 10% and 100% of the number of normal samples, and the total number of samples in the fine-tuning set is no less than 5000. During the pre-training phase, the physical flow encoder and semantic flow encoder are frozen, and only the anomaly detection head is fine-tuned. The anomaly detection head consists of a two-layer fully connected network and outputs a physically consistent perceived anomaly score. , , dimensionless, the larger the value, the higher the degree of anomaly.

[0089] The specific implementation method of step S04 is as follows.

[0090] Comprehensive Early Warning Index Score based on perceived physical consistency The proportion of anomalous energy in the graph-time joint time-frequency representation and the mean score of perceived physical consistency anomalies within the most recent 60 sampling points Calculated using the weighted summation method:

[0091] ;

[0092] In the formula, , , These are weighting coefficients, all dimensionless, ranging from 0.1 to 0.8, with a sum of 1. They were determined through a grid search experiment on the historical validation set, minimizing the sum of the false negative and false positive rates, with a step size of 0.05. for The maximum observed value in historical data, for The maximum observed value in historical data, for The maximum observed value in historical data, all three of which are dimensionless, is used for normalization to make all terms dimensionless. for The ratio of energy in the mid-abnormal frequency band to the total energy is dimensionless and ranges from 0 to 1. Within the last 60 sampling points The arithmetic mean of , dimensionless. This is a dimensionless comprehensive early warning index, ranging from 0 to 1. All terms in the equation are dimensionless and have uniform dimensions. Maintain normal monitoring at all times; when When a Level 1 warning is triggered, the cross-flow attention module is activated, and the sampling frequency is increased to twice the original sampling frequency; when When a level 2 early warning is triggered, a fault risk alert is sent and fault source node location is initiated; when When a Level 3 warning is triggered, data acquisition from the salinity sensor output channel and the dissolved oxygen sensor output channel is suspended, and resources are concentrated on continuous inference from the temperature sensor output channel, the pressure sensor output channel, and the power supply voltage acquisition channel.

[0093] The specific implementation method of step S05 is as follows.

[0094] For missing segments with a consecutive missing length exceeding 10 sampling points, a time-series missing data repair method based on a diffusion model is adopted, using the consistency residual of the temperature-salinity-density equation as a physical constraint to guide the denoising direction. For missing segments with a consecutive missing length not exceeding 10 sampling points, cubic spline interpolation is used. After repair, the telemetry time series is re-input into the dual-stream contrastive self-supervised transformer model to update the physical consistency sensing anomaly score. .

[0095] The specific implementation method of step S06 is as follows.

[0096] The knowledge distillation process uses a two-stream contrastive self-supervised transformer model as the teacher model and a lightweight convolutional network with no more than 2% of the parameters of the teacher model as the student model. The total training loss formula is as follows:

[0097] ;

[0098] In the formula, For the anomaly detection classification loss of the lightweight student model, dimensionless, The mean squared error distillation loss is the sum of the intermediate layer feature maps of the teacher model and the corresponding layer feature maps of the lightweight student model. It is dimensionless. , These are the corresponding weighting coefficients, all of which are dimensionless positive real numbers. The total training loss is dimensionless, with all terms in the equation being dimensionless and dimensionless. Integer 8-bit post-training quantization quantizes the floating-point weights and activation values ​​of the lightweight student model to an 8-bit integer representation. The quantization calibration dataset consists of at least 500 normally functioning samples. Structured pruning uses the output channels of each convolutional kernel... The norm is used for importance scoring. Channels with scores below a pruning threshold are removed. The pruning threshold is determined by limiting the decrease in accuracy of physical consistency perception anomalies to no more than 1%, and gradually increasing the pruning ratio until the model compression ratio is no less than 50 times. The compressed lightweight student model is deployed on an underwater embedded processor, utilizing ARM's official neural network inference library optimized for Cortex-M series processors. Single-instruction multiple-data-stream instructions are used to accelerate convolution and matrix multiplication operations in parallel, enabling real-time on-device warnings.

[0099] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effect of the invention, technicians set up a test environment and deployed a multi-parameter precision measuring instrument equipped with temperature, pressure, salinity, dissolved oxygen and power voltage sensors in a deep water area to collect online operation data for 210 consecutive days. The method of the invention was then used to perform full-process early warning processing.

[0100] During the data acquisition phase, the measuring instrument was deployed at a depth of approximately 1200 m, with a sampling frequency of 1 time / min, accumulating 302,400 sampling points. Among these, 23 fault periods were manually confirmed, covering three typical fault modes: zero-point drift of the pressure sensor, diaphragm aging of the dissolved oxygen sensor, and power supply voltage fluctuations. The data from the first 180 days was used as an unlabeled pre-training set, and the data from the following 30 days, including manually labeled fault periods, was used as a labeled fine-tuning set and test set.

[0101] In the Gaussian process regression drift compensation stage, temperature, pressure, and aging time were used as three-dimensional inputs to establish independent Gaussian process regression drift compensation models for each of the five sensors. The kernel function hyperparameters were iteratively updated on historical calibration data using the marginal log-likelihood maximization method. Taking the pressure sensor as an example, its aging time span covered 0–210 days. After compensation, the residual baseline drift in the sensor observation sequence was significantly reduced, such as… Figure 3 As shown, the comparison of sensor drift before and after compensation intuitively reflects the compensation effect of Gaussian process regression.

[0102] During the wavelet transform stage of the spectrogram, Granger causality tests were performed on the five sensors. The significance threshold was determined by multiple segmented statistical tests using historical normal operation data to ensure that the false alarm rate was below 2%. The test results constitute the directed edge weight matrix shown in Table 1, reflecting the causal strength between each sensor pair.

[0103] Table 1. Side weight matrix for sensor-Granger causality test

[0104] Based on the edge weight matrix in Table 1, a graph Laplacian matrix was constructed. After eigenvalue decomposition, a spectral wavelet basis was built. The compensated sensor observation sequence was then subjected to two-dimensional wavelet decomposition along the graph and time domains to obtain a graph-time joint time-frequency representation. In the pressure sensor drift fault that occurred on day 193, significant energy concentration was observed in the high-frequency spectral components at the pressure sensor node. Inverse solution of the thermonuclear equation located the fault source to the pressure sensor node, consistent with the results of manual confirmation.

[0105] During the training phase of the dual-stream contrastive self-supervised transformer model, the unlabeled pre-training set contained 180 days of normal operation data, sliced ​​into segments of 64 sampling points with a sliding window length, yielding approximately 270,000 samples. A conditional generative adversarial network was used to expand the fault samples in the labeled fine-tuning set from 23 segments to approximately 5200 samples, ensuring that the proportion of fault samples reached 15% of the normal samples, meeting the requirement that the total number of samples in the fine-tuning set be no less than 5000. In the pre-training phase, the physical flow encoder and semantic flow encoder were trained on the unlabeled pre-training set using the InfoNCE loss function, with a batch size of 128 and 200 training epochs. In the fine-tuning phase, the encoder was frozen, and the anomaly detection head was fine-tuned using the Focal Loss weighted loss function, with 50 training epochs. The dynamic routing trigger threshold was determined on the validation set by performing a grid search with the goal of maximizing fault recall, a search range of 0.5–0.95, a step size of 0.05, and was ultimately set at 0.75.

[0106] In the comprehensive early warning phase, the comprehensive early warning index is calculated by weighting three inputs with weights of 0.5, 0.3, and 0.2. These weights are determined using a grid search on the historical validation set, with the objective of minimizing the sum of the false negative and false positive rates. For example... Figure 2 As shown, the comprehensive early warning index began to rise continuously at approximately 30 sampling points before the fault occurred. On day 193, the comprehensive early warning index for the pressure sensor drift fault reached 0.63 approximately 29 minutes before the fault was confirmed, triggering a level-two early warning. A fault risk warning was sent to the shore-based system, and the fault source node location process was initiated. The location result was the pressure sensor node, which matched the manual confirmation result. The triggering status of each early warning level is shown in Table 2.

[0107] Table 2. Statistical Table of Triggering Each Warning Level

[0108] As shown in Table 2, the number of times the Level 3 warning was triggered was completely consistent with the number of confirmed fault segments, and there were no false alarms, indicating that the comprehensive warning index has strong accuracy in the high threshold range.

[0109] During the missing data repair phase, 17 consecutive missing segments appeared in the test set, of which 6 segments had a missing length exceeding 10 sampling points. These were repaired using a time-series missing data repair method based on a diffusion model, while the remaining 11 segments were processed using cubic spline interpolation. Taking the segment with 18 consecutive missing sampling points in the dissolved oxygen sensor channel on day 207 as an example, the mean residual of the consistency equation of the telemetry time series after repair was lower than that of the pure statistical method. This indicates that the guidance of physical constraints made the repair results more physically consistent. After the repaired data was input into the dual-stream comparison self-supervised transformer model, the fluctuation range of the physical consistency perception anomaly score was significantly narrowed, avoiding false alarms caused by repair errors.

[0110] During the embedded deployment phase, the dual-stream contrastive self-supervised transformer model, after knowledge distillation, yields a lightweight convolutional network student model with approximately 1.8% of the parameters of the teacher model. The INT8 quantization calibration dataset consists of 520 normal operation samples, significantly compressing the model weight storage after quantization. Structured pruning progressively increases the pruning ratio using L1 norm scoring, with the constraint that the accuracy of physical consistency perception anomaly scores does not decrease by more than 1%, ultimately achieving a compression ratio of approximately 52 times. The compressed model is deployed on an ARM Cortex-M7 architecture underwater embedded processor, utilizing SIMD instructions from the CMSIS-NN library to accelerate convolution and matrix multiplication operations, with single inference latency meeting real-time early warning requirements.

[0111] Compared to traditional single-sensor threshold judgment methods, this invention eliminates the interference of individualized aging drift on the fault judgment benchmark through Gaussian process regression drift compensation, making the compensated data have physically reliable reference values. By encoding the sensor network topology into the analysis framework through spectral wavelet transform, the propagation path of faults among multiple sensors can be tracked, thereby achieving node-level localization rather than relying solely on single-path threshold exceedance. Through a dual-stream contrastive self-supervised transformer model, the consistency between the original dynamic characteristics and physical equations is simultaneously perceived, enabling the learning of discriminative features even with a limited number of labeled samples, overcoming the inherent difficulty of scarce fault labeling data in deep-sea scenarios. The physical constraint repair mechanism of the diffusion model ensures the physical self-consistency of telemetry missing data segments, avoiding the contamination of subsequent anomaly detection processes by repair errors. The synergistic effect of these technical aspects forms a complete technical closed loop from sensor data compensation, topology perception, physical consistency judgment to missing data repair, substantially improving the overall early warning system's fault perception capability under complex deep-sea conditions.

[0112] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4.

[0113] Table 3. Variable Explanation Table (Part 1)

[0114] Table 4. Variable Explanation Table (Part Two)

[0115] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent early warning of operational faults in a marine multi-parameter precision measuring instrument, characterized in that, Includes the following steps: The time series of temperature, pressure, salinity, dissolved oxygen and power supply voltage of the multi-parameter measuring instrument are collected, input into the Gaussian process regression drift compensation model, and output the compensated sensor observation sequence of each sensor. The compensated sensor observation sequence is input into the spectrogram wavelet transform module. The edge weights are determined by the Granger causality test, the graph Laplacian matrix is ​​constructed, the graph-time joint time-frequency representation is output, and the fault source node is located in reverse. The residual sequences of the temperature-salinity-density equation and the dissolved oxygen saturation equation are calculated in real time and input together with the compensated sensor observation sequence into the dual-stream comparison self-supervised converter model to output the physical consistency perception anomaly score. The comprehensive early warning index is calculated by taking the physical consistency perception anomaly score, the proportion of anomaly energy in the graph-time joint time-frequency representation, and the mean of the physical consistency perception anomaly score in the most recent sampling point as inputs. The corresponding early warning level is triggered according to the interval to which the comprehensive early warning index belongs, and the inference calculation depth and sampling frequency are adjusted accordingly. For telemetry time series with missing segments, a time series missing data repair method based on diffusion model is used in combination with physical constraints to repair the data. The repaired telemetry time series is then input into a dual-stream comparison self-supervised transformer model to update the physical consistency perception anomaly score. The dual-stream contrastive self-supervised transformer model is compressed into a lightweight student model through knowledge distillation, quantization, and structured pruning. It is then deployed on an underwater embedded processor and the CMSIS-NN library is used to accelerate instruction-level inference on the ARM architecture, enabling real-time early warning at the device end.

2. The intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument according to claim 1, characterized in that, The Gaussian process regression drift compensation model takes temperature, pressure, and aging time as three-dimensional inputs and sensor baseline drift as the output target. It uses a squared exponential kernel function to construct the Gaussian process prior and uses historical online calibration data to iteratively update the kernel function hyperparameters. A set of model parameters is maintained independently for each sensor.

3. The intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument according to claim 2, characterized in that, The baseline drift of the sensors is the difference between the actual output value of each sensor under the current temperature, pressure, and aging time conditions and the factory calibration value, which is obtained by fitting historical online calibration data.

4. The intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument according to claim 3, characterized in that, The Granger causality test establishes univariate autoregressive models and bivariate autoregressive models for the sensor pairs respectively, compares the degree of decrease in prediction error, and establishes directed edges for sensor pairs whose decrease exceeds the significance threshold, using the decrease in prediction error as the edge weight of the graph Laplacian matrix.

5. The intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument according to claim 4, characterized in that, The graph Laplacian matrix is ​​obtained by subtracting the degree matrix and adjacency matrix of the weighted directed graph. After eigenvalue decomposition of the graph Laplacian matrix, a spectral wavelet basis is constructed. The compensated sensor observation sequence is subjected to two-dimensional wavelet decomposition along the graph domain and time domain to obtain the graph-time joint time-frequency representation.

6. The intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument according to claim 5, characterized in that, The dual-stream contrastive self-supervised transformer model includes a physical stream encoder, a semantic stream encoder, a dynamic routing module, a contrastive learning head, and an anomaly detection head. The physical stream encoder receives the compensated sensor observation sequence, and the semantic stream encoder receives the concatenated vector of the residual sequence of the temperature-salinity-density equation and the residual sequence of the dissolved oxygen saturation equation. Both encoders use rotational position encoding.

7. The intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument according to claim 6, characterized in that, The dynamic routing module calculates the cosine similarity between the physical flow feature vector and the semantic flow feature vector. When the cosine similarity is lower than the dynamic routing trigger threshold, the cross-flow attention module is activated to perform bidirectional interaction between the physical flow feature vector and the semantic flow feature vector.

8. The intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument according to claim 7, characterized in that, The contrastive learning head uses physical flow feature vectors and semantic flow feature vectors at the same time to form positive sample pairs, and feature vectors at other times to form negative samples. The InfoNCE loss function is used to guide the encoder to learn a compact representation of physical consistency awareness.

9. The intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument according to claim 8, characterized in that, The comprehensive early warning index is calculated by weighted summation of the physical consistency perception anomaly score, the proportion of anomaly energy in the graph-time joint time-frequency representation, and the mean of the physical consistency perception anomaly score within the most recent sampling point. The weights are determined by experiments on the historical validation set that minimize the sum of the false negative rate and the false positive rate.

10. The intelligent early warning method for operational faults of a marine multi-parameter precision measuring instrument according to claim 9, characterized in that, When the comprehensive early warning index is below the first early warning threshold, normal monitoring will continue. When the comprehensive early warning index is between the first and second early warning thresholds, a level one early warning is triggered, activating the cross-flow attention module and increasing the sampling frequency. When the comprehensive early warning index is between the second and third early warning thresholds, a level-two early warning is triggered, sending a fault risk warning to the shore-based system and initiating the fault source node location process. When the comprehensive early warning index is not lower than the third early warning threshold, a level 3 early warning is triggered, an emergency alarm is sent to the shore-based system, and data collection of non-critical parameter channels is suspended.

Images