State detection method for water energy storage unit

Abstract

The invention relates to the technical field of pumped storage, and discloses a state detection method for a water energy storage unit, which comprises the following steps: S1, deploying vibration, temperature, electrical and pressure sensors on the water energy storage unit, collecting multi-dimensional operation data, and carrying out noise reduction, standardization and time sequence alignment; s2, multi-dimensional features of the data are extracted through tensor decomposition, local space information is acquired by using a convolutional neural network, and a time sequence relation is modeled in combination with a Transform model; and S3, constructing a sensor topological structure through the graph convolutional network. The multi-source data fusion technology is adopted, data of vibration, temperature, electrical and pressure sensors are processed in a unified mode, time sequence prediction and an intelligent fault detection algorithm are combined, comprehensive health monitoring of the water energy storage unit is achieved, the detection precision is improved, meanwhile, potential faults can be found in time, and compared with data monitoring of a single sensor in a traditional method, the method has the advantage that the monitoring accuracy is improved. And the problem of inaccurate identification caused by large and complex data volume is obviously solved.

Description

Technical Field

[0001] The present invention relates to the technical field of pumped - storage energy, and specifically to a method for detecting the state of a water energy storage unit. Background Art

[0002] As an efficient energy storage and dispatching device, water energy storage units are widely used in power systems. Long - term stable operation depends on accurate state detection and fault early warning. Existing state detection methods for water energy storage units usually rely on single - type sensor data and judge whether a fault occurs based on fixed thresholds, which have many deficiencies in practical applications, resulting in limited effectiveness in the health management of the units.

[0003] Currently, the state monitoring of water energy storage units mainly analyzes single or a small amount of data from vibration, temperature, electrical, or pressure sensors. The operating state of the unit is the result of the interaction of multiple physical quantities. Relying on single - type data cannot comprehensively evaluate the health state, and at the same time, the data fusion ability of traditional methods is insufficient, resulting in a decrease in the accuracy of health assessment and prone to false alarms or omission of key faults.

[0004] The currently common method is to set a fixed threshold, and an alarm is triggered when the sensor data exceeds this threshold. The normal operating state of the unit varies under different working conditions, and the fixed threshold cannot adapt to dynamic changes. Therefore, the traditional fixed - threshold method has a high false - alarm rate or missed - alarm problems, affecting the reliability of detection.

[0005] The development of faults is progressive and gradually appears over a long time span. Traditional state detection methods usually analyze data based on short - term windows and cannot capture long - term trends. Therefore, existing methods cannot achieve true early warning.

[0006] Therefore, those skilled in the art provide a method for detecting the state of a water energy storage unit to solve the above - mentioned problems. Summary of the Invention

[0007] Aiming at the deficiencies of the prior art, the present invention provides a method for detecting the state of a water energy storage unit to solve the problems raised in the above background art.

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for detecting the state of a water energy storage unit includes:

[0009] S1, deploy vibration, temperature, electrical, and pressure sensors on the water energy storage unit, collect multi - dimensional operation data, and perform noise reduction, standardization, and time - series alignment;

[0010] S2. Extract multi-dimensional features of the data through tensor decomposition, obtain local spatial information using a convolutional neural network, and model the temporal relationship in combination with the Transformer model;

[0011] S3. Construct a sensor topology structure through a graph convolutional network, and complete fault detection in combination with the feature extraction results of S2;

[0012] S4. On the basis of fault detection, use a long short-term memory network to analyze short-term trends, use a Transformer network to extract long-term dependence features, judge the unit status in combination with a health index, and trigger an alarm when the health index is lower than the threshold;

[0013] S5. Perform preliminary processing on the warning data, calculate the probability distribution using Bayesian anomaly detection, and dynamically adjust the threshold in combination with the health index;

[0014] S6. Perform standardization processing on multi-sensor data, calculate the arithmetic mean to form a health index, and generate an alarm signal when the health index is lower than the set threshold;

[0015] S7. Collect and analyze the alarm signal, adjust the detection strategy according to the change of the health index, and update the model parameters.

[0016] Preferably, in step S1, the data acquisition and preprocessing further include:

[0017] Step 1.1. Deploy vibration sensors, temperature sensors, electrical sensors, and pressure sensors at key parts of the water energy storage unit to collect multi-dimensional operation data of the unit. The physical quantities of each sensor are defined as follows:

[0018] Vibration sensor: Collect the vibration acceleration A v , vibration velocity V v and vibration displacement D v ;

[0019] Temperature sensor: Collect the temperature T m of the surface of the unit components;

[0020] Electrical sensor: Collect the current I e , voltage U e and power P e of the motor;

[0021] Pressure sensor: Collect the pressure P h of the hydraulic system of the unit;

[0022] The sampling frequency of the sensor is set to f s , and a unified time step Δt is set;

[0023] Step 1.2. Remove noise using the wavelet transform method. The specific steps are as follows:

[0024] Let the original signal be X s (t), and perform decomposition using the discrete wavelet transform:

[0025] X s (t) = ∑ m ∑ n c m,n ψ m,n (t),

[0026] where c m,n is the wavelet coefficient, ψ m,n (t) is the wavelet basis function, m is the scale parameter, and n is the displacement parameter;

[0027] Set the wavelet threshold T w :

[0028] where σ w is the standard deviation of the wavelet coefficient, and N is the data length;

[0029] Retain the low-frequency signal and reconstruct the signal after noise reduction:

[0030]

[0031] where c ′ m,n is the wavelet coefficient after threshold processing, and is the signal after reconstructing and noise reduction;

[0032] Step 1.3. Standardize the data. Let the data of a certain sensor be X d , and the standardization formula is as follows:

[0033] where X ′ d is the standardized data, μ d is the mean of X d :

[0034] σ d is the standard deviation of X d :

[0035] where X d,i is the data value of the i-th sample in dimension d;

[0036] Step 1.4. Since the sampling frequencies f s of different sensors are different, it is necessary to perform time series alignment processing;

[0037] Suppose there is a sampling frequency \(f\) of low-frequency data low and a sampling frequency \(f\) of high-frequency data high of sensor data;

[0038] Suppose \(f\) high = \(kf\) low , and interpolate the low-frequency data to align it with the time step of the high-frequency data:

[0039] Adopt the linear interpolation method:

[0040]

[0041] where \(X\) low (\(t_1\)) and \(X\) low (\(t_2\)) are adjacent low-frequency data points, \(t_1\) and \(t_2\) are timestamps, and \(t\) is the time point to be interpolated;

[0042] After interpolation, all data is aligned to the time step \(\Delta t\);

[0043] Step 1.5, store the processed data, construct a unified multi-dimensional data format, and define the data matrix \(X\) final :

[0044]

[0045] where each row represents the data of a time step, each column represents a sensor variable, and the total number of rows \(N\) depends on the sampling duration;

[0046] After data formatting, enter Step S2, where the data will be used for tensor decomposition and feature extraction.

[0047] Preferably, in the said Step S2, the multi-source data fusion and fault detection further includes:

[0048] Step 2.1, in Step S1, after sensor deployment, data denoising, normalization processing, and time series alignment, obtain the multi-dimensional data matrix \(X\) final ;

[0049] Adopt the tensor decomposition method for feature extraction, and the specific steps are as follows:

[0050] Represent the processed multi-dimensional sensor data as a high-order tensor:

[0051] where \(I\) is the number of sensors, \(J\) is the length of the time series, \(K\) is the feature dimension at each time point, is the tensor;

[0052] Adopt the CP decomposition method for the tensor Decompose to extract potential features and patterns in the data. The mathematical expression of tensor decomposition is as follows:

[0053]

[0054] where R is the rank of tensor decomposition, A r is the first - dimensional matrix of the r - th mode, B r is the second - dimensional matrix of the r - th mode, C r is the third - dimensional matrix of the r - th mode, and ° is the tensor product operation;

[0055] The optimization objective of tensor decomposition is to minimize the decomposition error, that is:

[0056]

[0057] where ∥·∥ 2 F is the Frobenius norm;

[0058] Step 2.2, in Step S1, the data after data denoising, normalization, and time - series alignment is ready to enter the fault detection stage;

[0059] Adopt a convolutional neural network, which is used to extract local features;

[0060] Set the input data X ∈ R I×J×K , where I is the number of sensors, J is the length of the time series, and K is the feature dimension at each time point;

[0061] The convolutional neural network performs a convolution operation on the input data to extract spatial features. The mathematical expression of the convolution operation is as follows:

[0062] h = f(W * X + b),

[0063] where h is the feature map output by the convolutional layer, W is the convolution kernel, b is the bias term, f is the activation function, * is the convolution operation, and X is the input data;

[0064] Step 2.3, in Step 2.2, the convolutional neural network extracts local spatial features, and there are long - term dependence relationships in the local spatial features. To capture long - term dependencies, a Transformer model is used for time - series data modeling;

[0065] The Transformer model captures long - term dependence relationships in time - series data through the self - attention mechanism. The mathematical expression of the self - attention operation is as follows:

[0066]

[0067] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector, softmax(·) is the normalization operation, and T is the transpose operation of the matrix;

[0068] Step 2.4: Combine the local spatial features extracted by the convolutional neural network and the long-term temporal relationship modeled by the Transformer model to construct an overall fault detection model. The decision logic for fault detection is as follows:

[0069] Based on the above-extracted features, the fault detection model, on the basis of multi-source data fusion, combines the outputs of the convolutional neural network and the Transformer model to determine whether there is a fault in the unit;

[0070] If a fault is detected, the system will trigger an alarm signal through the early warning mechanism, further calculate the health index, and enter step S3 for status analysis.

[0071] Preferably, in step S3, the sensor topology modeling and fault detection based on the graph convolutional network further include:

[0072] In step S2, extract the multi-dimensional features of the data through the tensor decomposition method, obtain the local spatial information using the convolutional neural network, and establish the temporal dependence relationship in combination with the Transformer model;

[0073] In step S3, further construct the sensor topology structure and use the graph convolutional network for fault detection;

[0074] Step 3.1: In step 2.1, the high-order tensor of the sensor data has been constructed and the key features are extracted through tensor decomposition. Now, use the features to construct a sensor topology graph so that the system can perform fault detection based on the physical correlation between sensors;

[0075] Step 3.1.1: Represent the sensor system of the water energy storage unit as a weighted undirected graph:

[0076] G = (V, E, W), where V is the set of sensors, E is the connection relationship between sensors, and W is the adjacency matrix;

[0077] Let W be the weight matrix, and each element w in the matrix ij defines the correlation weight between sensor i and sensor j, and uses the Gaussian kernel function to calculate the weight:

[0078]

[0079] where f i 、f i represent the feature vectors extracted by sensors i and j, ||fi -f j where || is the Euclidean distance and σ is the scale parameter;

[0080] Step 3.1.2 is for graph convolution calculation. Define the normalized Laplacian matrix L:

[0081] L = D - W,

[0082] where L is the normalized Laplacian matrix, D is the degree matrix, and W is the adjacency matrix;

[0083] Adopt the normalized Laplacian matrix:

[0084]

[0085] where I is the identity matrix;

[0086] Step 3.2: In Step 2.3, the Transformer model has been used to capture the temporal dependencies, while the graph convolutional network is used to capture the topological correlations between sensors;

[0087] Step 3.2.1: The calculation formula of graph convolution is as follows:

[0088]

[0089] where H (l) is the node feature matrix of the l-th layer, is the normalized Laplacian matrix, W (l) is the learning weight matrix of the l-th layer, b (l) is the bias term, and σ(·) is the activation function;

[0090] Step 3.2.2: Adopt the graph convolutional network structure and calculate as follows:

[0091] The first layer:

[0092] The output layer:

[0093] Finally, output H (2) as the fault state probabilities of each sensor;

[0094] Step 3.3: Fault classification and decision-making:

[0095] S3.3.1: Use the high-dimensional features H (2) extracted by the graph convolutional network and adopt a fully connected neural network for fault classification:

[0096] y = softmax(W f H (2) + b f ),

[0097] Among them, W f is the fault classification weight matrix, b f is the bias term, and y is the probability distribution of each fault category;

[0098] Step 3.3.2, set the threshold T f . If the probability of a certain category is greater than T f , it is determined that the unit has a fault of this category:

[0099]

[0100] If the maximum probability is less than T f , it is considered that the unit is in a normal state;

[0101] Step 3.4, if a fault is detected, calculate the health index, further analyze the short-term trend and long-term dependence relationship by combining LSTM and Transformer, and generate a fault report;

[0102] If no fault is detected, continue to monitor the state of the water energy storage unit and wait for subsequent data input.

[0103] Preferably, in the step S4, the fault trend analysis and warning trigger based on the health index further include:

[0104] In step S3, the sensor topology structure has been modeled by the graph convolutional network, and the fault detection is combined with multi-dimensional features to output the fault classification result. However, there will be short-term fluctuations in single fault detection. Therefore, in S4, it is necessary to further analyze the development trend of the fault, extract the short-term trend and long-term dependence features by combining the long short-term memory network and Transformer, and calculate the health index. When the health index is lower than the set threshold, a warning is triggered;

[0105] Step 4.1, in step 3.3, the fault classification probability y of each sensor has been calculated by the graph convolutional network. Next, it is necessary to calculate the health index based on the fault classification result;

[0106] The health index is used to quantify the operating state of the water energy storage unit, and its value range is usually [0, 1]. A low value indicates a poor health condition of the equipment;

[0107] Step 4.1.1, define the health index HI of each sensor i :

[0108]

[0109] Among them, p i,c is the fault probability of sensor i in fault category c, w cis the severity weight of fault category c, and C is the number of fault categories;

[0110] For the health index HI of the overall unit sys , take the weighted average of the health indices of all sensors:

[0111]

[0112] where HI sys is the health index, v i is the importance coefficient of sensor i, and I is the total number of sensors;

[0113] Step 4.1.2, to reduce the influence of short-term fluctuations on the health index, use the exponential weighted moving average to smooth the health index:

[0114]

[0115] where is the health index at the current moment, is the health index at the previous moment, and α is the smoothing coefficient;

[0116] Thus, the smoothed health index is obtained;

[0117] Step 4.2, in Step 3.4, the fault state of the unit has been obtained and the health index has been calculated. To predict the short-term trend of the health index, use the long short-term memory network for modeling;

[0118] Step 4.2.1, the long short-term memory network is suitable for modeling the short-term dynamic changes of time series, and the core calculation formula is as follows:

[0119] Forget gate:

[0120] Input gate:

[0121] Candidate state:

[0122] Cell state update:

[0123] Output gate:

[0124] Hidden state update: h t = o t ·tanh(C t )

[0125] where f t 、i t 、o t are the forget gate, input gate and output gate, and C tis the cell state of the long short-term memory network, h t is the hidden state of the long short-term memory network, W f 、W i 、W C 、W o are the trainable weights of the long short-term memory network, b f 、b i 、b C 、b o are the bias parameters;

[0126] The h calculated by the long short-term memory network t is used as the basis for predicting the short-term health index :

[0127]

[0128] Among them, W h is the weight matrix for linearly mapping the hidden state at time t, b h represents the bias vector added to the result of the linear mapping;

[0129] Step 4.3, in Step 2.3, the Transformer has been used to extract long-term dependencies to further extract the long-term trend of the health index using the Transformer network;

[0130] Step 4.3.1, since the Transformer does not have temporal order information, positional encoding needs to be added:

[0131]

[0132] Among them, pos is the time step, i is the position index, d is the Transformer dimension,

[0133] PE (pos,2i) represents the positional encoding value when the dimension index is even, and PE (pos,2i+1) represents the positional encoding value when the dimension index is odd;

[0134] Step 4.3.2, the Transformer uses the multi-head self-attention mechanism to calculate the long-term trend:

[0135]

[0136] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d k is the dimension of the key vector, softmax(·) is the normalization operation, and T is the transpose operation of the matrix;

[0137] Finally, use the feature H output by the TransformerT Calculating long-term health index prediction:

[0138]

[0139] where, represents the unit health index predicted after pushing back T time steps from the current time t, H T represents the feature vector extracted from the multi-dimensional sensor data at time T, W T represents the weight matrix that maps the feature vector H T to the health index prediction space, b T represents the bias vector added to the result of the linear mapping;

[0140] Step 4.4, based on the short-term and long-term health index predictions, set the warning threshold H T :

[0141] trigger warning,

[0142] If the predicted health index drops below the threshold, trigger a warning for anomaly detection and threshold adjustment;

[0143] If the health index is below the warning threshold, perform Bayesian anomaly detection, analyze whether there is a false alarm, and dynamically adjust the threshold to generate a warning report;

[0144] If the health index is normal, continue to monitor the health index and keep the status updated.

[0145] Preferably, in the step S5, the Bayesian anomaly detection and threshold dynamic adjustment further include:

[0146] Step 5.1, Bayesian anomaly detection and probability distribution calculation:

[0147] In step S4, a warning signal has been triggered and the prediction of the health index has been obtained In this step, first perform anomaly detection on the health index, use the Bayesian method to evaluate the anomaly probability of the health index, judge whether there is a false alarm, and dynamically adjust the threshold. The specific steps are as follows:

[0148] Step 5.1.1, construction of the anomaly detection model:

[0149] According to the historical health index data and the current predicted value Establish a Bayesian model of Gaussian distribution, assuming that the health index conforms to Gaussian distribution N(μ,σ 2 ), where μ is the mean and σ 2 is the variance;

[0150] First, calculate the mean and variance of the current health index:

[0151]

[0152] where k is the time step of historical data, is the health index at time step t - i;

[0153] Step 5.1.2, Abnormality detection probability calculation:

[0154] After determining the Gaussian distribution model of the health index, use Bayes' theorem to calculate whether the current health index prediction is abnormal; first calculate the posterior probability of the current health index value:

[0155]

[0156] where, is the probability density of the health index under the given mean μ and variance σ 2 condition, is the predicted health index value;

[0157] Step 5.1.3, Abnormality threshold determination:

[0158] According to the calculated posterior probability value set the abnormality threshold τ ab , if the posterior probability value of the current health index is less than this threshold, it is determined as abnormal and enters the abnormality handling process;

[0159] trigger anomaly detection process,

[0160] where τ ab is the preset anomaly detection threshold.

[0161] Preferably, in step 5.2, the threshold is dynamically adjusted:

[0162] To ensure that the fault detection system has an adaptive ability, this step includes the dynamic adjustment of the health index threshold. According to the Bayesian anomaly detection results and the changes in real-time data, the preset health index threshold is adjusted. The specific steps are as follows:

[0163] Step 5.2.1, Threshold adjustment algorithm:

[0164] According to the current health index and the anomaly value calculated from the Bayesian posterior probability, dynamically adjust the threshold T H to adapt to the current health state of the system. Set the new threshold T H ′ The calculation formula is as follows:

[0165]

[0166] Among them, T H is the originally set health index threshold, α is the adjustment coefficient, is the currently predicted health index, and μ is the mean of historical data;

[0167] Step 5.3, result output and entering the next step:

[0168] If an anomaly is detected, trigger the anomaly handling process, generate a health index alarm signal and perform alarm analysis, and continue to monitor the system health status according to the adjusted threshold to dynamically adjust the detection strategy;

[0169] If no anomaly is detected, continuously monitor the health index, maintain the current threshold, and enter the next round of health status analysis.

[0170] Preferably, in step S6, the health index standardization and alarm signal generation further include:

[0171] In step S5, through Bayesian anomaly detection and dynamic threshold adjustment, the anomaly situation of the health index has been judged, and the health index threshold is adjusted as needed. Next, in step S6, the comprehensive health index will be calculated based on the standardized data of multiple groups of sensors to generate an alarm signal;

[0172] Step 6.1, standardization processing of the health index:

[0173] In steps S4 and S5, the health indices of each sensor and the predicted health index of the unit have been obtained. To make the data of different sensors comparable, first perform standardization processing on the health indices of each sensor. The specific steps are as follows:

[0174] Step 6.1.1, health index standardization:

[0175] By standardizing the health index of each sensor, for the health index HI of sensor i i , the standardization formula is:

[0176]

[0177] Among them, HI i is the health index of sensor i, μ i is the mean of the health index of sensor i, σ i is the standard deviation of the health index of sensor i, is the standardized health index;

[0178] Step 6.2, calculation of the comprehensive health index:

[0179] In steps S4 and S5, a separate health index is calculated for each sensor, and the health index is adjusted by the Bayesian method. Next, the comprehensive health index v of the overall unit is calculated based on the standardized health indices of the sensors. i The specific steps are as follows:

[0180] Step 6.2.1, calculating the comprehensive health index by the weighted average method:

[0181] The comprehensive health index HI sys is calculated based on the health index and importance weight of each sensor. Assuming that the importance coefficient of sensor i is v i , the calculation formula for the health index of the overall unit is:

[0182]

[0183] where is the standardized health index of sensor i, v i is the importance coefficient of sensor i, I is the total number of sensors, and HI sys is the health index of the overall unit;

[0184] Step 6.3, generating an alarm signal:

[0185] The judgment basis for generating an alarm signal is as follows:

[0186] if HI sys < T H , trigger an alarm signal,

[0187] where HI sys is the health index of the unit at the current moment, and T H is the set alarm threshold;

[0188] If the health index is lower than the set threshold, trigger an alarm signal, collect and analyze the alarm signal, and adjust the detection strategy and model parameters according to the change of the health index;

[0189] If the health index is normal, continue to monitor the health index and maintain the current detection strategy and model.

[0190] A terminal device, the terminal device includes:

[0191] At least one set of sensors for collecting the operation data of the water energy storage unit;

[0192] A data processing unit for denoising, standardizing, and time series alignment of the collected multi-dimensional data;

[0193] A health index calculation module for calculating the health index of a water energy storage unit based on the processed data;

[0194] An anomaly detection module for determining whether to trigger an alarm signal based on health index calculation and Bayesian anomaly detection algorithm;

[0195] A display module for displaying alarm signals and health status information;

[0196] A control module for adjusting detection strategies and model parameters when anomalies are detected.

[0197] A storage medium stores computer-executable instructions, which when executed by a terminal device, implement the following steps:

[0198] Collect the operation data of the water energy storage unit through at least one set of sensors;

[0199] Perform noise reduction, normalization, and time series alignment on the data;

[0200] Calculate the health index and determine whether to trigger an alarm signal;

[0201] If the health index is lower than the set threshold, trigger an alarm signal and enter the alarm signal collection and analysis process;

[0202] Adjust the detection strategy and update the model parameters according to the alarm signal;

[0203] Through a feedback adjustment mechanism, the health status detection process of the unit is optimized in real time.

[0204] The present invention provides a method for detecting the state of a water energy storage unit. It has the following beneficial effects:

[0205] 1. The present invention adopts a multi-source data fusion technology to uniformly process the data of vibration, temperature, electrical, and pressure sensors, combines time series prediction and intelligent fault detection algorithms to achieve all-round health monitoring of the water energy storage unit, improves the detection accuracy, and can timely discover potential faults. Compared with the data monitoring of a single sensor in the traditional method, it significantly solves the problem of inaccurate identification caused by large and complex data.

[0206] 2. The present invention introduces time series modeling based on Transformer and sensor topology structure analysis of graph convolutional network, enabling fault diagnosis to dynamically adapt to different states of the unit operation, effectively capturing long-term dependence relationships in the fault detection process, avoiding false alarms and missed alarms compared with the traditional method of using fixed thresholds and simple models, and improving the stability of the unit.

[0207] 3. The health index calculation and feedback mechanism of the present invention adjusts the threshold through Bayesian anomaly detection, making the alarm sensitive and adaptive. Compared with the traditional method that relies on manual threshold setting in the prior art, the present invention can adjust the threshold according to real-time data changes, greatly improving the accuracy and timeliness of fault warning, ensuring that the unit can be promptly responded to when a fault occurs in the early stage, and reducing downtime and losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0208] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0209] To enable those skilled in the art to understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0210] The present invention will be described in detail below with reference to the accompanying drawings:

[0211] Embodiment:

[0212] Please refer to the attached Figure 1 , the embodiment of the present invention provides a method for detecting the state of a water energy storage unit, including:

[0213] S1. Deploy vibration, temperature, electrical, and pressure sensors on the water energy storage unit, collect multi-dimensional operation data, and perform noise reduction, standardization, and time series alignment;

[0214] Step 1.1. Deploy vibration sensors, temperature sensors, electrical sensors, and pressure sensors at key parts of the water energy storage unit to collect multi-dimensional operation data of the unit. The physical quantities of each sensor are defined as follows:

[0215] Vibration sensor: Collect the vibration acceleration A v , vibration velocity V v and vibration displacement D v ;

[0216] Temperature sensor: Collect the temperature T m on the surface of the unit components;

[0217] Electrical sensor: Collect the current I e , voltage U e and power P e of the motor;

[0218] Pressure sensor: Collect the pressure P h of the hydraulic system of the unit;

[0219] The sampling frequency of the sensor is set to f s , and a unified time step Δt is set;

[0220] Step 1.2, adopt the wavelet transform method to remove noise, and the specific steps are as follows:

[0221] Let the original signal be X s (t), and perform decomposition using discrete wavelet transform:

[0222] X s (t) = ∑ m ∑ n c m,n ψ m,n (t),

[0223] where c m,n is the wavelet coefficient, ψ m,n (t) is the wavelet basis function, m is the scale parameter, and n is the displacement parameter;

[0224] Set the wavelet threshold T w :

[0225] where σ w is the standard deviation of the wavelet coefficient, and N is the data length;

[0226] Retain the low-frequency signal Reconstruct the signal after noise reduction:

[0227]

[0228] where c ′ m,n is the wavelet coefficient after threshold processing, is the signal after reconstructing and noise reduction;

[0229] Step 1.3, perform standardization processing on the data. Let the data of a certain sensor be X d , and the standardization formula is as follows:

[0230] where X ′ d is the data after standardization, μ d is the mean of X d :

[0231] σ d is the standard deviation of X d :

[0232] where X d,iis the data value of the i-th sample in dimension d;

[0233] Step 1.4, due to the different sampling frequencies f of different sensors s being different, temporal alignment processing is required;

[0234] Suppose there is sensor data with a sampling frequency f of low-frequency data low and a sampling frequency f of high-frequency data high ;

[0235] Suppose f high = kf low , and interpolate the low-frequency data to align it to the time step of the high-frequency data:

[0236] Adopt the linear interpolation method:

[0237]

[0238] where X low (t1) and X low (t2) are adjacent low-frequency data points, t1 and t2 are timestamps, and t is the time point to be interpolated;

[0239] After interpolation is completed, all data is aligned to the time step Δt;

[0240] Step 1.5, store the processed data, construct a unified multi-dimensional data format, and define the data matrix X final :

[0241]

[0242] where each row represents the data of a time step, each column represents a sensor variable, and the total number of rows N depends on the sampling duration;

[0243] After the data is formatted, enter step S2, where the data will be used for tensor decomposition and feature extraction;

[0244] S2, extract the multi-dimensional features of the data through tensor decomposition, obtain local spatial information using a convolutional neural network, and model the temporal relationship by combining with the Transformer model;

[0245] Step 2.1, in step S1, after sensor deployment, data denoising, normalization processing, and temporal alignment, obtain the multi-dimensional data matrix X final ;

[0246] Adopt the tensor decomposition method for feature extraction, and the specific steps are as follows:

[0247] Represent the processed multi-dimensional sensor data as a high-order tensor:

[0248] Where \(I\) is the number of sensors, \(J\) is the length of the time series, and \(K\) is the feature dimension at each time point, is a tensor;

[0249] The CP decomposition method is used to decompose the tensor to extract the latent features and patterns in the data. The mathematical expression of the tensor decomposition is as follows:

[0250]

[0251] Where \(R\) is the rank of the tensor decomposition, \(A\) r is the first-dimensional matrix of the \(r\)-th mode, \(B\) r is the second-dimensional matrix of the \(r\)-th mode, \(C\) r is the third-dimensional matrix of the \(r\)-th mode, and \(\circ\) is the tensor product operation;

[0252] The optimization objective of the tensor decomposition is to minimize the decomposition error, that is:

[0253]

[0254] Where \(\|\cdot\|\) 2 F is the Frobenius norm;

[0255] Step 2.2, in Step S1, the data after data denoising, normalization, and time series alignment is ready to enter the fault detection stage;

[0256] A convolutional neural network is used, and the convolutional neural network is used to extract local features;

[0257] Set the input data \(X\in\mathbb{R}\) I×J×K , where \(I\) is the number of sensors, \(J\) is the length of the time series, and \(K\) is the feature dimension at each time point;

[0258] The convolutional neural network performs a convolution operation on the input data to extract spatial features. The mathematical expression of the convolution operation is as follows:

[0259] \(h = f(W*X + b)\),

[0260] where \(h\) is the feature map output by the convolutional layer, \(W\) is the convolutional kernel, \(b\) is the bias term, \(f\) is the activation function, \(*\) is the convolution operation, and \(X\) is the input data;

[0261] Step 2.3, in Step 2.2, the convolutional neural network extracts local spatial features, and there are long-term dependency relationships in the local spatial features. To capture long-term dependencies, a Transformer model is used for time series data modeling;

[0262] The Transformer model captures long-term dependencies in time series data through the self-attention mechanism. The mathematical expression of the self-attention operation is as follows:

[0263]

[0264] where Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector, softmax(·) is the normalization operation, and T is the matrix transpose operation;

[0265] Step 2.4: Combine the local spatial features extracted by the convolutional neural network and the long-term time series relationship modeled by the Transformer model to construct an overall fault detection model. The decision logic for fault detection is as follows:

[0266] Based on the above-extracted features, the fault detection model, on the basis of multi-source data fusion, combines the outputs of the convolutional neural network and the Transformer model to determine whether there is a fault in the unit;

[0267] When a fault is detected, the system will trigger an alarm signal through the warning mechanism, further calculate the health index, and enter step S3 for status analysis;

[0268] S3: Construct a sensor topology structure through the graph convolutional network, and combine the feature extraction results of S2 to complete fault detection;

[0269] Step 3.1: In step 2.1, a high-order tensor of sensor data has been constructed and key features have been extracted through tensor decomposition. Now, use the features to construct a sensor topology graph so that the system can perform fault detection based on the physical correlation between sensors;

[0270] Step 3.1.1: Represent the sensor system of the water energy storage unit as a weighted undirected graph:

[0271] G = (V, E, W), where V is the set of sensors, E is the connection relationship between sensors, and W is the adjacency matrix;

[0272] Let W be the weight matrix, and each element w in the matrix ij defines the correlation weight between sensor i and sensor j, and the Gaussian kernel function is used to calculate the weight:

[0273]

[0274] where f i and f i represent the feature vectors extracted by sensors i and j, ||f i - f j || is the Euclidean distance, and σ is the scale parameter;

[0275] Step 3.1.2, for graph convolution calculation, define the normalized Laplacian matrix L:

[0276] L = D - W,

[0277] where L is the normalized Laplacian matrix, D is the degree matrix, and W is the adjacency matrix;

[0278] Adopt the normalized Laplacian matrix:

[0279]

[0280] where I is the identity matrix;

[0281] Step 3.2, in Step 2.3, the time-dependent relationship has been captured by the Transformer model, while the graph convolutional network is used to capture the topological association between sensors;

[0282] Step 3.2.1, the calculation formula of graph convolution is as follows:

[0283]

[0284] where H (l) is the node feature matrix of the l-th layer, is the normalized Laplacian matrix, W (l) is the learning weight matrix of the l-th layer, b (l) is the bias term, and σ(·) is the activation function;

[0285] Step 3.2.2, adopt the graph convolutional network structure, and the calculation is as follows:

[0286] The first layer:

[0287] The output layer:

[0288] Finally, output H (2) as the fault state probability of each sensor;

[0289] Step 3.3, fault classification and decision-making:

[0290] S3.3.1, use the high-dimensional feature H (2) extracted by the graph convolutional network, and adopt a fully connected neural network for fault classification:

[0291] y = softmax(W f H (2) + b f ),

[0292] where W f is the fault classification weight matrix, bf is the bias term, and y is the probability distribution of each fault category;

[0293] Step 3.3.2, set the threshold T f , if the probability of a certain category is greater than T f , it is judged that the unit has a fault of this category:

[0294]

[0295] If the maximum probability is less than T f , it is considered that the unit is in a normal state;

[0296] Step 3.4, if a fault is detected, calculate the health index, and further analyze the short-term trend and long-term dependence relationship by combining LSTM and Transformer to generate a fault report;

[0297] If no fault is detected, continue to monitor the state of the water energy storage unit and wait for subsequent data input;

[0298] S4, on the basis of fault detection, use the long short-term memory network to analyze the short-term trend, use the Transformer network to extract long-term dependence features, combine the health index to judge the state of the unit, and trigger an early warning when the health index is lower than the threshold;

[0299] Step 4.1, in Step 3.3, the fault classification probability y of each sensor has been calculated through the graph convolutional network. Next, it is necessary to calculate the health index based on the fault classification results;

[0300] The health index is used to quantify the operating state of the water energy storage unit, and its value range is usually [0, 1]. A low value indicates a poor health condition of the equipment;

[0301] Step 4.1.1, define the health index HI of each sensor i :

[0302]

[0303] where p i,c is the fault probability of sensor i in fault category c, w c is the severity weight of fault category c, and C is the number of fault categories;

[0304] For the health index HI of the overall unit sys , take the weighted average of the health indexes of all sensors:

[0305]

[0306] where HI sys is the health index, and vi is the importance coefficient of sensor i, and I is the total number of sensors;

[0307] Step 4.1.2. To reduce the impact of short-term fluctuations on the health index, exponential weighted moving average is used to smooth the health index:

[0308]

[0309] where is the health index at the current moment, is the health index at the previous moment, and α is the smoothing coefficient;

[0310] Thus, the smoothed health index is obtained;

[0311] Step 4.2. In Step 3.4, the fault state of the unit has been obtained and the health index has been calculated. To predict the short-term trend of the health index, a long short-term memory network is used for modeling;

[0312] Step 4.2.1. The long short-term memory network is suitable for modeling the short-term dynamic changes of time series. The core calculation formula is as follows:

[0313] Forget gate:

[0314] Input gate:

[0315] Candidate state:

[0316] Cell state update:

[0317] Output gate:

[0318] Hidden state update: h t = o t ·tanh(C t )

[0319] where f t 、i t 、o t are the forget gate, input gate and output gate, C t is the cell state of the long short-term memory network, h t is the hidden state of the long short-term memory network, W f 、W i 、W C 、W o are the trainable weights of the long short-term memory network, b f 、b i 、b C 、b o are the bias parameters;

[0320] The h calculated by the long short - term memory network t As the basis for short - term health index prediction :

[0321]

[0322] Among them, W h is the weight matrix for linearly mapping the hidden state at time t, and b h represents the bias vector added to the result of the linear mapping;

[0323] Step 4.3. In step 2.3, the Transformer has been used to extract long - term dependencies to further extract the long - term trend of the health index using the Transformer network;

[0324] Step 4.3.1. Since the Transformer does not have time - order information, position encoding needs to be added:

[0325]

[0326] Among them, pos is the time step, i is the position index, d is the dimension of the Transformer,

[0327] PE (pos,2i) represents the position - encoding value when the dimension index is even, and PE (pos,2i+1) represents the position - encoding value when the dimension index is odd;

[0328] Step 4.3.2. The Transformer uses the multi - head self - attention mechanism to calculate the long - term trend:

[0329]

[0330] Among them, Q is the query matrix, K is the key matrix, V is the value matrix, d k is the dimension of the key vector, softmax(·) is the normalization operation, and T is the transpose operation of the matrix;

[0331] Finally, use the feature H output by the Transformer T to calculate the long - term health index prediction:

[0332]

[0333] Among them, represents the unit health index predicted after pushing T time steps backward from the current time t, and H T represents the feature vector extracted from multi - dimensional sensor data at time T, and W T represents the feature vector HT The weight matrix mapped to the health index prediction space, b T represents the bias vector added to the linear mapping result;

[0334] Step 4.4, based on the short-term and long-term health index predictions, set the warning threshold H T :

[0335] trigger warning,

[0336] If the predicted health index drops below the threshold, trigger a warning for anomaly detection and threshold adjustment;

[0337] If the health index is below the warning threshold, perform Bayesian anomaly detection, analyze whether there are false alarms, and dynamically adjust the threshold to generate a warning report;

[0338] If the health index is normal, continue to monitor the health index and keep the status updated;

[0339] S5, perform preliminary processing on the warning data, calculate the probability distribution using Bayesian anomaly detection, and dynamically adjust the threshold in combination with the health index;

[0340] Step 5.1, Bayesian anomaly detection and probability distribution calculation:

[0341] In step S4, a warning signal has been triggered and the prediction of the health index has been obtained In this step, first perform anomaly detection on the health index, use the Bayesian method to evaluate the anomaly probability of the health index, judge whether there are false alarms, and dynamically adjust the threshold. The specific steps are as follows:

[0342] Step 5.1.1, construction of the anomaly detection model:

[0343] Based on the historical health index data and the current predicted value Establish a Bayesian model with a Gaussian distribution, assuming that the health index conforms to a Gaussian distribution N(μ,σ 2 ) under normal conditions, where μ is the mean and σ 2 is the variance;

[0344] First calculate the mean and variance of the current health index:

[0345]

[0346] where k is the time step of the historical data, is the health index at time step t-i;

[0347] Step 5.1.2, anomaly detection probability calculation:

[0348] After determining the Gaussian distribution model of the health index, use Bayes' theorem to calculate the prediction of the current health index whether it is abnormal; first calculate the posterior probability of the current health index value:

[0349]

[0350] where is the probability density of the health index under the given mean μ and variance σ 2 condition, is the predicted health index value;

[0351] Step 5.1.3, anomaly threshold determination:

[0352] According to the calculated posterior probability value set the anomaly threshold τ ab , if the posterior probability value of the current health index is less than this threshold, it is determined to be abnormal and enter the anomaly handling process;

[0353] trigger anomaly detection process,

[0354] where τ ab is the preset anomaly detection threshold;

[0355] Step 5.2, dynamic threshold adjustment:

[0356] To ensure that the fault detection system has an adaptive ability, this step includes the dynamic adjustment of the health index threshold. According to the Bayesian anomaly detection results and the changes in real-time data, the preset health index threshold is adjusted. The specific steps are as follows:

[0357] Step 5.2.1, threshold adjustment algorithm:

[0358] According to the current health index and the anomaly value calculated from the Bayesian posterior probability, dynamically adjust the threshold T H to adapt to the current health state of the system. Set the new threshold T H ′ The calculation formula of H is as follows:

[0359]

[0360] where T H is the originally set health index threshold, α is the adjustment coefficient, is the currently predicted health index, μ is the mean of historical data;

[0361] Step 5.3, result output and enter the next step:

[0362] If an anomaly is detected, trigger the anomaly handling process, generate a health index alarm signal and conduct alarm analysis, and continue to monitor the system health status according to the adjusted threshold to dynamically adjust the detection strategy;

[0363] If no anomaly is detected, continuously monitor the health index, maintain the current threshold, and enter the next round of health status analysis;

[0364] S6. Standardize the multi-sensor data, calculate the arithmetic mean to form a health index, and generate an alarm signal when the health index is lower than the set threshold;

[0365] Step 6.1. Standardization of the health index:

[0366] In steps S4 and S5, the health indices of each sensor and the predicted health index of the unit have been obtained. To make the data of different sensors comparable, first standardize the health indices of each sensor. The specific steps are as follows:

[0367] Step 6.1.1. Standardization of the health index:

[0368] By standardizing the health indices of each sensor, for the health index HI of sensor i i , the standardization formula is:

[0369]

[0370] where HI i is the health index of sensor i, μ i is the mean of the health index of sensor i, σ i is the standard deviation of the health index of sensor i, is the standardized health index;

[0371] Step 6.2. Calculation of the comprehensive health index:

[0372] In steps S4 and S5, a separate health index is calculated for each sensor, and the health index is adjusted by the Bayesian method. Next, calculate the comprehensive health index v of the overall unit according to the standardized health indices of each sensor i , and the specific steps are as follows:

[0373] Step 6.2.1. Calculation of the comprehensive health index by the weighted average method:

[0374] The comprehensive health index HI sys is calculated based on the health indices and importance weights of each sensor. Assuming that the importance coefficient of sensor i is v i , the formula for calculating the health index of the overall unit is:

[0375]

[0376] Among them, is the health index of sensor i after standardization, v i is the importance coefficient of sensor i, I is the total number of sensors, HI sys is the health index of the overall unit;

[0377] Step 6.3, alarm signal generation:

[0378] The judgment basis for alarm signal generation is as follows:

[0379] if HI sys < T H , trigger the alarm signal,

[0380] Among them, HI sys is the health index of the unit at the current moment, T H is the set alarm threshold;

[0381] If the health index is lower than the set threshold, trigger the alarm signal, collect and analyze the alarm signal, and adjust the detection strategy and model parameters according to the change of the health index;

[0382] If the health index is normal, continue to monitor the health index and maintain the current detection strategy and model;

[0383] S7, collect and analyze the alarm signal, adjust the detection strategy according to the change of the health index, and update the model parameters.

[0384] The sensors deployed in step S1 can cover multiple physical quantities of the unit, avoiding information loss caused by a single sensor; after data noise reduction processing, the data is clear and false alarms are reduced; the benefit of standardizing the data can unify the data range, prevent a certain type of data from affecting the analysis result, and improve the calculation stability and optimize the subsequent modeling effect; through the linear interpolation method, align the low-frequency data to the high-frequency data time step, ensure the integrity of the data without loss, ensure the consistency of subsequent analysis, and solve the out-of-step problem to avoid time misalignment affecting the model judgment; convert the preprocessed data into a unified format and construct a time-sensor matrix, which can reduce the storage overhead;

[0385] In step S2, tensor decomposition constructs sensor data into a high-order tensor, uses CP decomposition to extract features, realizes data dimensionality reduction, reduces redundant information, improves computational efficiency, and extracts potential patterns to find hidden operating rules; local feature extraction uses a convolutional neural network to extract local features of the unit state, realizes automatic learning of important features without relying on manual selection, enhances local fault recognition ability, and captures subtle abnormal changes; the convolutional network in long-term dependence modeling can capture short-term features, while the Transformer can extract long-term dependence relationships, can detect gradually deteriorating faults in advance, avoid the limitations of simple threshold methods, and reduce the false alarm rate;

[0386] Fault detection decision-making constructs a fault detection model, conducts multi-dimensional analysis, reduces misjudgment, dynamically adapts to different operating states, and improves detection accuracy;

[0387] In step S3, sensor topology structure modeling uses a graph convolutional network to model the relationships between sensors, accurately evaluates the fault influence range, reduces false alarms caused by local anomalies, and improves stability; information transmission between sensors uses a graph convolutional network to calculate the features of sensors, and combines the temporal features of the Transformer for global correlation analysis to detect systematic problems, and then can comprehensively judge the fault type and distinguish local anomalies from overall trend changes; health index calculation calculates the overall health index of the unit according to the fault probabilities of each sensor. A low value indicates a poor health condition;

[0388] In step S4, short-term prediction uses a long short-term memory network to predict short-term trends to adapt to sudden faults, can quickly respond to health deterioration, and reduces short-term noise interference to provide smooth prediction; long-term trend analysis uses the Transformer to extract long-term dependence features, avoids the impact of sudden anomalies on overall judgment, and arranges maintenance in advance; dynamic warning sets a dynamic threshold according to the short-term and long-term prediction results. When the health index is lower than the threshold, a warning is triggered, and Bayesian anomaly detection is combined to avoid false alarms;

[0389] In step S5, Bayesian anomaly detection calculates the probability distribution of the health index based on historical data to detect whether the current state is abnormal. If the probability is less than the threshold, it is determined to be abnormal to reduce false alarms and improve detection accuracy; threshold adaptive adjustment can dynamically adjust the threshold to adapt to different operating conditions and reduce the uncertainty brought by manual setting;

[0390] In step S6, data standardization ensures the comparability of data from different sensors and avoids the influence of numerical scale differences on the results; calculate the comprehensive health index using weighted average to reflect the overall unit state; alarm logic: when the health index is lower than the set threshold, an alarm is triggered, and the system continuously monitors and dynamically adjusts the strategy;

[0391] In step S7, the early warning data analysis is used to analyze the alarm signal, determine whether it is a false alarm, and optimize the decision-making model by combining the historical health index data; the model is updated, the fault detection parameters are adjusted regularly to adapt to the equipment aging trend, and the algorithm is optimized by combining the actual fault conditions to continuously improve the detection ability.

[0392] In summary, the present invention adopts multi-source data fusion, intelligent fault detection, and adaptive early warning mechanism to realize the comprehensive health management of the water energy storage unit.

[0393] A terminal device, the terminal device includes:

[0394] At least one group of sensors for collecting the operation data of the water energy storage unit;

[0395] A data processing unit for denoising, normalizing, and time series alignment of the collected multi-dimensional data;

[0396] A health index calculation module for calculating the health index of the water energy storage unit according to the processed data;

[0397] An anomaly detection module for determining whether to trigger an alarm signal based on the health index calculation and Bayesian anomaly detection algorithm;

[0398] A display module for displaying the alarm signal and health status information;

[0399] A control module for adjusting the detection strategy and model parameters when an anomaly is detected.

[0400] A storage medium stores computer-executable instructions, and when the instructions are executed by the terminal device, the following steps are implemented:

[0401] Collect the operation data of the water energy storage unit through at least one group of sensors;

[0402] Perform denoising, normalizing, and time series alignment on the data;

[0403] Calculate the health index and determine whether to trigger an alarm signal;

[0404] If the health index is lower than the set threshold, trigger an alarm signal and enter the alarm signal collection and analysis process;

[0405] Adjust the detection strategy and update the model parameters according to the alarm signal;

[0406] Through the feedback adjustment mechanism, the health status detection process of the unit is optimized in real time.

[0407] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the state of a water energy storage unit, characterized in that, Including: S1. Deploy vibration, temperature, electrical, and pressure sensors on the water energy storage unit to collect multi-dimensional operation data, and perform noise reduction, standardization, and time series alignment; S2. Extract multi-dimensional features of the data through tensor decomposition, use a convolutional neural network to obtain local spatial information, and combine with the Transformer model to model the time series relationship; S3. Construct a sensor topology structure through a graph convolutional network, and combine with the feature extraction results of S2 to complete fault detection; S4. On the basis of fault detection, use a long short-term memory network to analyze short-term trends, use a Transformer network to extract long-term dependence features, combine with a health index to judge the unit status, and trigger an alarm when the health index is lower than the threshold; S5. Perform preliminary processing on the warning data, use Bayesian anomaly detection to calculate the probability distribution, and dynamically adjust the threshold in combination with the health index; S6. Perform standardization processing on the multi-sensor data, calculate the arithmetic mean to form a health index, and generate an alarm signal when the health index is lower than the set threshold; S7. Collect and analyze the alarm signal, adjust the detection strategy according to the change of the health index, and update the model parameters.

2. The state detection method of a water energy storage unit according to claim 1, characterized in that In the step S1, the data collection and preprocessing further includes: Step 1.

1. Deploy vibration sensors, temperature sensors, electrical sensors, and pressure sensors at key parts of the water energy storage unit to collect multi-dimensional operation data of the unit. The physical quantities of each sensor are defined as follows: Vibration sensor: collects the vibration acceleration A of the unit v , vibration speed V v and vibration displacement D v ; Temperature sensor: Collect the temperature T of the surface of the unit components m ; Electrical sensor: Collect the current I of the motor e , voltage U e and power P e ; Pressure sensor: Collect the pressure P of the hydraulic system of the unit h ; The sampling frequency of the sensor is set to f s , and a unified time step Δt is set; Step 1.

2. Use the wavelet transform method to remove noise. The specific steps are as follows: Let the original signal be X s (t), and perform decomposition using discrete wavelet transform: X s (t) = ∑ m ∑ n c m,n ψ m,n (t), where c m,n is the wavelet coefficient, ψ m,n (t) is the wavelet basis function, m is the scale parameter, and n is the displacement parameter; Set the wavelet threshold T w : where σ w is the standard deviation of the wavelet coefficients, and N is the data length; Retain low-frequency signals Reconstruct the signal after noise reduction: Among them, c′ m,n is the wavelet coefficient after threshold processing, and is the signal after reconstruction and noise reduction; Step 1.3, perform normalization processing on the data. Let the data of a certain sensor be X d , and the normalization formula is as follows: Among them, X' d is the standardized data, and μ d is the mean value of X d : σ d is X d standard deviation of: where X d,i is the data value of the i-th sample in dimension d; Step 1.

4. Since the sampling frequencies f of different sensors s are different, it is necessary to perform timing alignment processing; Suppose there is sensor data with a sampling frequency \(f\) of low-frequency data low and a sampling frequency \(f\) of high-frequency data high ; Assume f high = kf low , interpolate and align the low-frequency data to the time step of the high-frequency data: Adopt the linear interpolation method: Among them, X low (t1) and X low (t2) are adjacent low-frequency data points, where t1 and t2 are timestamps, and t is the time point for interpolation; After interpolation, all data are aligned to the time step Δt; Step 1.5, store the processed data, construct a unified multi-dimensional data format, and define the data matrix X final : Among them, each row represents the data of a time step, each column represents a sensor variable, and the total number of rows N depends on the sampling duration; After data formatting, enter step S2, where the data will be used for tensor decomposition and feature extraction.

3. The state detection method of a water energy storage unit according to claim 1, characterized in that In the step S2, the multi-source data fusion and fault detection further includes: Step 2.1, in Step S1, after sensor deployment, data denoising, normalization, and time series alignment, a multi-dimensional data matrix X is obtained final ; Use the tensor decomposition method for feature extraction. The specific steps are as follows: Represent the processed multi-dimensional sensor data as a high-order tensor: where I is the number of sensors, J is the length of the time series, and K is the feature dimension at each time point, is a tensor; Use the CP decomposition method to decompose the tensor to extract the latent features and patterns in the data. The mathematical expression of tensor decomposition is as follows: where R is the rank of tensor decomposition, A r is the first-dimensional matrix of the r-th mode, B r is the second-dimensional matrix of the r-th mode, C r is the third-dimensional matrix of the r-th mode, and ° is the tensor product operation; The optimization objective of tensor decomposition is to minimize the decomposition error, that is: wherein, is the Frobenius norm; Step 2.

2. In step S1, the data after noise reduction, standardization, and time series alignment are ready to enter the fault detection stage; Use a convolutional neural network, and the convolutional neural network is used to extract local features; Set the input data \(X\in\mathbb{R}^{I\times J\times K}\), where \(I\) is the number of sensors, \(J\) is the length of the time series, and \(K\) is the feature dimension at each time point; I×J×K ​ The convolutional neural network performs a convolution operation on the input data to extract spatial features. The mathematical expression of the convolution operation is as follows: h = f(W * X + b), where h is the feature map output by the convolutional layer, W is the convolutional kernel, b is the bias term, f is the activation function, * is the convolution operation, and X is the input data; Step 2.

3. In step 2.2, the convolutional neural network extracts local spatial features, and there are long-term dependence relationships in the local spatial features. To capture long-term dependencies, a Transformer model is used for time series data modeling; The Transformer model captures the long-term dependence relationship in the time series data through the self-attention mechanism. The mathematical expression of the self-attention operation is as follows: Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector, softmax(·) is the normalization operation, and T is the transpose operation of the matrix; Step 2.4: Combine the local spatial features extracted by the convolutional neural network and the long-term temporal relationship modeled by the Transformer model to construct an overall fault detection model. The decision logic for fault detection is as follows: Based on the above-extracted features, the fault detection model, on the basis of multi-source data fusion, combines the outputs of the convolutional neural network and the Transformer model to determine whether there is a fault in the unit; When a fault is detected, the system will trigger an alarm signal through the warning mechanism, further calculate the health index, and enter step S3 for status analysis.

4. A method for detecting the state of a water energy storage unit according to claim 1, characterized in that In the above step S3, the sensor topology modeling and fault detection based on the graph convolutional network further include: In step S2, multi-dimensional features of the data are extracted by the tensor decomposition method, local spatial information is obtained by the convolutional neural network, and the temporal dependence relationship is established in combination with the Transformer model; In step S3, it is necessary to further construct the sensor topology structure and use the graph convolutional network for fault detection; Step 3.1: In step 2.1, the high-order tensor χ of the sensor data has been constructed, and key features are extracted by tensor decomposition. Now, use the features to construct a sensor topology graph so that the system can perform fault detection based on the physical correlation between sensors; Step 3.1.1: Represent the sensor system of the water energy storage unit as a weighted undirected graph: G=(V, E, W), where V is the set of sensors, E is the connection relationship between sensors, and W is the adjacency matrix; Let \(W\) be the weight matrix, and each element \(w\) in the matrix ij defines the correlation weight between sensor \(i\) and sensor \(j\), and the Gaussian kernel function is used to calculate the weight: where, f i , f i represent the feature vectors extracted by sensors i and j, ||f i - f j || is the Euclidean distance, and σ is the scale parameter; Step 3.1.2: For graph convolution calculation, define the normalized Laplacian matrix L: L = D - W, where L is the normalized Laplacian matrix, D is the degree matrix, and W is the adjacency matrix; Adopt the normalized Laplacian matrix: where I is the identity matrix; Step 3.2: In step 2.3, the time dependence relationship has been captured by the Transformer model, and the graph convolutional network is used to capture the topological correlation between sensors; Step 3.2.1: The calculation formula of graph convolution is as follows: Among them, H (l) is the node feature matrix of the l-th layer, is the normalized Laplacian matrix, W (l) is the learning weight matrix of the l-th layer, b (l) is the bias term, and σ(·) is the activation function; Step 3.2.2: Adopt the graph convolutional network structure, and the calculation is as follows: The first layer: Output layer: Finally, output H (2) As the probability of the failure state of each sensor; Step 3.3: Fault classification and decision-making: S3.3.1, the high-dimensional feature H extracted using the graph convolutional network (2) , and a fully-connected neural network is used for fault classification: y = softmax(W f H (2) + b f ), Among them, W f is the fault classification weight matrix, b f is the bias term, and y is the probability distribution of each fault category; Step 3.3.2, set the threshold value T f , if the probability of a certain category is greater than T f , it is determined that the unit has a fault of this category: If the maximum probability is less than T f , it is considered that the unit is in a normal state; Step 3.4: If a fault is detected, calculate the health index, further analyze the short-term trend and long-term dependence relationship in combination with LSTM and Transformer, and generate a fault report; If no fault is detected, continue to monitor the status of the water energy storage unit and wait for subsequent data input.

5. The state detection method of a water energy storage unit according to claim 1, characterized in that In the above step S4, the fault trend analysis and warning trigger based on the health index further include: In step S3, the sensor topology structure has been modeled by the graph convolutional network, and fault detection is performed in combination with multi-dimensional features, and the fault classification result is output. However, there will be short-term fluctuations in a single fault detection. Therefore, in S4, it is necessary to further analyze the development trend of the fault, extract short-term trend and long-term dependence features in combination with the long short-term memory network and Transformer, and calculate the health index. When the health index is lower than the set threshold, a warning is triggered; Step 4.1: In step 3.3, the fault classification probability y of each sensor has been calculated by the graph convolutional network. Next, it is necessary to calculate the health index based on the fault classification result; The health index is used to quantify the operating state of the water energy storage unit, usually with a value range of [0, 1]. A low value indicates poor equipment health; Step 4.1.1, define the health index HI of each sensor i : where p i,c is the failure probability of sensor i for failure category c, w c is the severity weight of failure category c, and C is the number of failure categories; For the health index HI of the overall unit sys , take the weighted average of the health indices of all sensors: Among them, HI sys is the health index, v i is the importance coefficient of sensor i, and I is the total number of sensors; Step 4.1.2, to reduce the impact of short-term fluctuations on the health index, exponential weighted moving average is used to smooth the health index: Among them, is the health index at the current moment, is the health index at the previous moment, and α is the smoothing coefficient; Thus, the smoothed health index is obtained; Step 4.2, in Step 3.4, the fault state of the unit has been obtained and the health index has been calculated. To predict the short-term trend of the health index, a long short-term memory network is used for modeling; Step 4.2.1, the long short-term memory network is suitable for modeling the short-term dynamic changes of time series. The core calculation formula is as follows: Forgotten Gate: Input gate: Candidate status: Cell status update: Output gate: Hidden state update: h t = o t ·tanh(C t ) where, f t , i t , o t are the forget gate, input gate, and output gate, C t is the cell state of the long short-term memory network, h t is the hidden state of the long short-term memory network, W f , W i , W C , W o are the trainable weights of the long short-term memory network, b f , b i , b C , b o are the bias parameters; h calculated by the long short-term memory network t As the basis for short-term health index prediction : Among them, W h is the weight matrix for linearly mapping the hidden state at time t, and b h represents the bias vector added to the result of the linear mapping; Step 4.3, in Step 2.3, the Transformer has been used to extract long-term dependencies. To further extract the long-term trend of the health index, the Transformer network is adopted; Step 4.3.1, since the Transformer does not have time sequence information, position encoding needs to be added: where pos is the time step, i is the position index, and d is the Transformer dimension, PE (pos,2i) represents the position encoding value when the dimension index is even, and PE (pos,2i+1) represents the position encoding value when the dimension index is odd; Step 4.3.2, the Transformer uses the multi-head self-attention mechanism to calculate the long-term trend: Among them, Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector, softmax(·) is the normalization operation, and T is the transpose operation of the matrix; Finally, use the feature H output by the Transformer T Calculate the long-term health index prediction: Among them, represents the predicted unit health index after pushing back T time steps from the current moment t, H T represents the feature vector extracted from the multi-dimensional sensor data at time T, W T represents the feature vector H T The weight matrix that maps to the health index prediction space, b T represents the bias vector added to the result of the linear mapping; Step 4.4, set an early warning threshold H based on the short-term and long-term health index predictions T : If the predicted health index drops below the threshold, a warning is triggered for anomaly detection and threshold adjustment; If the health index is lower than the warning threshold, Bayesian anomaly detection is performed to analyze whether there is a false alarm, and the threshold is dynamically adjusted to generate a warning report; If the health index is normal, continue to monitor the health index and keep the status updated.

6. The state detection method of a water energy storage unit according to claim 1, characterized in that, In the said Step S5, the Bayesian anomaly detection and threshold dynamic adjustment further include: Step 5.1, Bayesian anomaly detection and probability distribution calculation: In step S4, a warning signal has been triggered, and a prediction of the health index has been obtained. In this step, anomaly detection is first performed on the health index. The Bayesian method is used to evaluate the anomaly probability of the health index, determine whether there is a false alarm, and dynamically adjust the threshold. The specific steps are as follows: Step 5.1.1, construction of the anomaly detection model: Based on historical health index data and current predicted values A Bayesian model with a Gaussian distribution is established, assuming that the health index conforms to the Gaussian distribution N(μ,σ 2 ) under normal circumstances, where μ is the mean and σ 2 is the variance; First, calculate the mean and variance of the current health index: where k is the time step of historical data, is the health index at time step t - i; Step 5.1.2, calculation of the anomaly detection probability: After determining the Gaussian distribution model of the health index, use Bayes' theorem to calculate the prediction of the current health index whether it is abnormal; first calculate the posterior probability of the current health index value: Among them, is the probability density of the health index under the given mean μ and variance σ 2 conditions, is the predicted health index value; Step 5.1.3, determination of the anomaly threshold: According to the calculated posterior probability value Set the anomaly threshold τ ab , if the posterior probability value of the current health index is less than this threshold, it is determined as an anomaly and enters the anomaly handling process; Among them, τ ab is a preset anomaly detection threshold.

7. A method for detecting the state of a water energy storage unit according to claim 6, characterized in that The said Step 5.2, dynamic adjustment of the threshold: To ensure that the fault detection system has an adaptive ability, this step includes the dynamic adjustment of the health index threshold. Based on the results of Bayesian anomaly detection and according to the changes in real-time data, the preset health index threshold is adjusted. The specific steps are as follows: Step 5.2.1, threshold adjustment algorithm: According to the current health index and the outliers calculated by Bayesian posterior probability, dynamically adjust the threshold T H to adapt to the current health state of the system. Set the new threshold T′ H The calculation formula is as follows: where, T H is the originally set health index threshold, α is the adjustment coefficient, is the currently predicted health index, and μ is the mean of historical data; Step 5.3, result output and entry into the next step: If an anomaly is detected, trigger the anomaly handling process to generate a health index alarm signal and perform alarm analysis, and continue to monitor the system health status according to the adjusted threshold to dynamically adjust the detection strategy; If no anomaly is detected, continuously monitor the health index, maintain the current threshold, and enter the next round of health status analysis.

8. The state detection method of a water energy storage unit according to claim 1, characterized in that, In the said Step S6, the health index standardization and alarm signal generation further include: In Step S5, through Bayesian anomaly detection and dynamic adjustment of the threshold, the anomaly situation of the health index has been judged, and the health index threshold has been adjusted as needed. Next, in Step S6, the comprehensive health index will be calculated based on the standardized data of multiple groups of sensors to generate an alarm signal; Step 6.1, standardization processing of the health index: In steps S4 and S5, the health indices of each sensor and the predicted health index of the unit have been obtained. To make the data of different sensors comparable, the health indices of each sensor are first standardized. The specific steps are as follows: Step 6.1.1, Health index standardization: By standardizing the health index of each sensor, for the health index HI of sensor i i , the standardization formula is: Among them, HI i is the health index of sensor i, μ i is the mean of the health index of sensor i, σ i is the standard deviation of the health index of sensor i, is the standardized health index; Step 6.2, Comprehensive health index calculation: In steps S4 and S5, a separate health index is calculated for each sensor, and the health index is adjusted by the Bayesian method. Next, the comprehensive health index v of the entire unit is calculated based on the standardized health indices of the respective sensors i , and the specific steps are as follows: Step 6.2.1, Calculate the comprehensive health index by the weighted average method: Comprehensive Health Index HI sys It is calculated based on the health indices and importance weights of each sensor. Assuming the importance coefficient of sensor i is v i , then the calculation formula for the health index of the overall unit is as follows: Among them, is the health index of sensor i after standardization, v i is the importance coefficient of sensor i, I is the total number of sensors, HI sys is the health index of the whole unit; Step 6.3, Alarm signal generation: The judgment basis for alarm signal generation is as follows: ifHI sys <T H ,trigger alarm signal, Among them, HI sys is the unit health index at the current moment, and T H is the set alarm threshold value; If the health index is lower than the set threshold, trigger an alarm signal, collect and analyze the alarm signal, and adjust the detection strategy and model parameters according to the change of the health index; If the health index is normal, continue to monitor the health index and maintain the current detection strategy and model.

9. A terminal device, characterized in that, The terminal device includes: At least one set of sensors for collecting the operation data of the water energy storage unit; A data processing unit for denoising, standardizing and time series alignment of the collected multi-dimensional data; A health index calculation module for calculating the health index of the water energy storage unit according to the processed data; An anomaly detection module for judging whether to trigger an alarm signal based on the health index calculation and the Bayesian anomaly detection algorithm; A display module for displaying the alarm signal and health status information; A control module for adjusting the detection strategy and model parameters when an anomaly is detected.

10. A storage medium, characterized in that, Stored with computer-executable instructions, when the instructions are executed by the terminal device, the following steps are implemented: Collect the operation data of the water energy storage unit through at least one set of sensors; Denoise, standardize and time series align the data; Calculate the health index and judge whether to trigger an alarm signal; If the health index is lower than the set threshold, trigger an alarm signal and enter the alarm signal collection and analysis process; Adjust the detection strategy and update the model parameters according to the alarm signal; Through the feedback adjustment mechanism, optimize the health status detection process of the unit in real time.

Images