Energy storage liquid cooling system fault prediction method and device

Abstract

The present invention provides a method and device for predicting faults in an energy storage liquid cooling system, relating to the fields of artificial intelligence and data processing technology. By dynamically analyzing the local noise level and time-scale variations in monitoring data and adaptively segmenting the raw data, the method can better capture the characteristics of different fault modes. It also dynamically adjusts the denoising strategy based on the characteristics of each segmented data, effectively removing noise while maximizing the retention of useful information. By precisely controlling the denoising process, the present invention ensures that even in high-noise environments, subtle but critical fault characteristics can be accurately extracted, which is crucial for improving the sensitivity and specificity of fault diagnosis systems.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and data processing technology, and in particular to a method and device for predicting faults in an energy storage liquid cooling system. Background Art

[0002] Energy storage liquid cooling systems play a vital role in areas such as battery storage and power dispatching. Their stability and safety directly affect the efficient operation of the power system. With the continuous development of energy storage technology, the complexity of liquid cooling systems and the diversity of application environments have placed higher requirements on system monitoring and fault diagnosis. Key equipment involved in energy storage liquid cooling systems, such as coolant pipes, battery modules, and circulation pumps, are usually equipped with multiple types of sensors to monitor important parameters such as temperature, pressure, and flow in real time. However, these system monitoring data are often interfered with by high-frequency noise, nonlinear time series characteristics, and weak fault characteristics. Traditional fault diagnosis methods often have difficulty processing such complex time-varying signals, and the diagnostic accuracy is insufficient under multiple fault modes.

[0003] Currently, many liquid cooling system fault prediction methods rely on simple statistical feature extraction or fixed-threshold denoising methods. These methods are often unable to adapt to the differences in noise intensity and time scales under different failure modes of liquid cooling systems, resulting in insufficient extraction of fault features. Furthermore, existing machine learning models often face the problem of being unable to effectively capture temporal dependencies and long-term dependency features when processing multivariate coupled and time-varying data. Furthermore, when new failure modes emerge, traditional models need to be fully retrained, resulting in high computational overhead and time delays.

[0004] Therefore, improving the accuracy of liquid cooling system fault prediction, especially accurately capturing the characteristics of different types of faults in complex environments, has become a technical challenge that needs to be solved urgently. Existing methods often ignore the time-varying and multi-scale characteristics of data, making it difficult for the models to effectively address the common sudden and slow-changing fault characteristics during the actual operation of energy storage liquid cooling systems. Furthermore, the fixed threshold denoising and manual reliance on data labeling also limit the robustness and adaptability of existing technologies. Summary of the Invention

[0005] In view of this, an object of the present invention is to provide a method and device for predicting faults in an energy storage liquid cooling system to solve the above-mentioned problems.

[0006] In a first aspect, an embodiment of the present invention provides a method for predicting faults of an energy storage liquid cooling system, wherein the method includes: obtaining operation monitoring parameters corresponding to a preset energy storage liquid cooling system; the operation monitoring parameters include monitoring parameters corresponding to preset key equipment of the energy storage liquid cooling system; based on the local noise level and time scale changes of the operation monitoring parameters, data of the operation monitoring parameters are divided to obtain segmented data corresponding to multiple time periods; according to the noise intensity corresponding to the segmented data, a denoising strength corresponding to the segmented data is determined; based on the denoising strength, the segmented data is denoised to determine the parameters to be measured corresponding to each time scale of the operation monitoring parameters; the parameters to be measured are input into a pre-trained neural network model, and the parameters to be measured are classified and predicted by the neural network model to determine the fault category prediction results corresponding to the parameters to be measured; according to the fault category prediction results, the fault state corresponding to the energy storage liquid cooling system is determined.

[0007] In combination with the first embodiment of the first aspect, the steps of classifying and predicting the parameters to be measured through a neural network model and determining the fault category prediction results corresponding to the parameters to be measured include: using the time series feature dynamic fusion layer of the neural network model, using multiple convolution kernels of different scales to capture the features of the parameters to be measured, and determining the local target features corresponding to the parameters to be measured; using the adaptive regression unit of the neural network model, using a preset expert selection matrix to adaptively select the activation path of the parameters to be measured, and determining the high-dimensional feature representation corresponding to the parameters to be measured; based on the preset step-size convolution operation, displaying the time scale features corresponding to the local target features and the high-dimensional feature representation respectively; based on the contribution of the parameters to be measured to the time series feature dynamic fusion layer or the adaptive regression unit, adaptively fusing the time scale features corresponding to the local target features and the time scale features corresponding to the high-dimensional feature representation, and determining the target fusion features corresponding to the parameters to be measured; performing fault category probability prediction on the target fusion features, and determining the fault category prediction results corresponding to the parameters to be measured.

[0008] In combination with the second embodiment of the first aspect, the steps of capturing features of the parameters to be measured and determining the local target features corresponding to the parameters to be measured by using a plurality of convolution kernels of different scales through the dynamic fusion layer of the temporal features of the neural network model include: performing nonlinear mapping on the feature parameters of the parameters to be measured at each moment through a preset multi-layer perceptron to determine the query vectors corresponding to the parameters to be measured at each moment; determining the convolution kernel attention weights corresponding to the feature parameters of the parameters to be measured at each moment according to the attention mechanism of the query vector and the key vector of the preset dynamic convolution kernel; capturing the temporal features of the parameters to be measured and determining the local target features corresponding to the parameters to be measured according to the convolution kernel attention weights and the time attention weights corresponding to the feature parameters; wherein the time attention weights are determined based on the extraction of local context information of the parameters to be measured using a time sliding window.

[0009] In combination with the third embodiment of the first aspect, the adaptive regression unit of the neural network model uses a preset expert selection matrix to adaptively select the activation path of the parameter to be measured, and the step of determining the high-dimensional feature representation corresponding to the parameter to be measured includes: using the preset expert selection matrix to map the parameter to be measured into an expert probability distribution; based on the expert probability distribution, determining the expert units corresponding to the characteristic parameters of the parameter to be measured at each moment; based on the regression parameters corresponding to the expert units, performing a nonlinear transformation on the parameter to be measured to determine the high-dimensional feature representation corresponding to the parameter to be measured.

[0010] In combination with the fourth embodiment of the first aspect, the operation monitoring parameters are divided into data based on the local noise level and time scale change of the operation monitoring parameters to obtain segmented data corresponding to multiple time scales, including: determining the data noise difference in the current time period according to the data noise level corresponding to each time point of the preset time period of the operation monitoring parameters; determining the time segmentation point corresponding to the operation monitoring parameters based on the data noise difference; dividing the operation monitoring parameters into data based on the time segmentation point and the preset time segment length penalty item to obtain segmented data corresponding to multiple time scales.

[0011] In combination with the fifth implementation of the first aspect, the method further includes: obtaining a pre-constructed training sample set; inputting the training sample set into a preset neural network model in batches, and calculating the classification error, regression error and total loss corresponding to the expert diversity constraint item of the training sample set through a preset hybrid loss function; updating the parameters of the neural network model according to the total loss; and, during the parameter update process of the neural network model, based on a preset time-sensitive discarding strategy, probabilistically randomly discarding the feature elements of the training sample set; the probability of random discarding is calculated based on the feature amplitude and diversity corresponding to the feature elements; until the neural network model meets the preset training requirements, the current optimal weight parameters are selected to generate the target neural network model.

[0012] In combination with the sixth implementation manner of the first aspect, the method for calculating the loss function includes: determining the weighted regression loss weight according to the number of samples corresponding to different sample categories of the preset training sample set; performing parameter space differentiation constraints on the parameter vectors of the preset multiple expert units according to the preset parameter vector constraint algorithm, and determining the expert diversity constraint loss corresponding to the training sample set; wherein the expert unit is used to learn the preset target fault mode characteristics in the training sample set based on the corresponding parameter space; and calculating the hybrid loss function corresponding to the training sample set according to the expert diversity constraint loss, the weighted regression loss weight and the preset classification loss.

[0013] In combination with the seventh implementation method of the first aspect, the step of updating the parameters of the neural network model according to the total loss includes: obtaining the model parameters corresponding to the neural network model in each iteration, and constraining the parameter update amplitude of the neural network model based on the parameter state corresponding to the model parameters; extracting the total loss corresponding to the pre-stored memory samples from a pre-set memory library, and updating the parameters of the neural network model according to the total loss corresponding to the memory samples; wherein the memory samples are samples in the training sample set.

[0014] In combination with the eighth implementation of the first aspect, the method further includes: determining new samples corresponding to the training sample set through online learning; measuring the similarity between the new samples and the training sample set, and determining the difference samples between the new samples and the training sample set; and updating the memory bank based on the difference samples.

[0015] In a second aspect, an embodiment of the present invention further provides a fault prediction device for an energy storage liquid cooling system, wherein the device includes: a data acquisition module for acquiring operation monitoring parameters corresponding to a preset energy storage liquid cooling system; the operation monitoring parameters include monitoring parameters corresponding to preset key equipment of the energy storage liquid cooling system; a data processing module for dividing the operation monitoring parameters based on the local noise level and time scale change of the operation monitoring parameters, and obtaining segmented data corresponding to multiple time periods; a calculation module for determining the denoising strength corresponding to the segmented data according to the noise intensity corresponding to the segmented data; a denoising module for denoising the segmented data based on the denoising strength, and determining the parameters to be measured corresponding to each time scale of the operation monitoring parameters; an execution module for inputting the parameters to be measured into a pre-trained neural network model, classifying and predicting the parameters to be measured through the neural network model, and determining the fault category prediction results corresponding to the parameters to be measured; and an output module for determining the fault state corresponding to the energy storage liquid cooling system according to the fault category prediction results.

[0016] The embodiments of the present invention bring about the following beneficial effects: The embodiments of the present invention provide a method and device for predicting faults in an energy storage liquid cooling system, which realizes adaptive segmentation processing of the original data by dynamically analyzing the local noise level and time scale changes of the monitoring data. It can better capture the characteristics of different fault modes, especially faults with significant time scale differences (such as sudden faults and slow-changing faults). In addition, the embodiments of the present invention also dynamically adjust the denoising strategy according to the characteristics of each segmented data, which can not only effectively remove noise but also retain useful information to the maximum extent, which helps to improve the data quality of subsequent model training. The embodiments of the present invention ensure that even in a high-noise environment, weak but critical fault features can be accurately extracted by precisely controlling the denoising process, which is crucial for improving the sensitivity and specificity of the fault diagnosis system.

[0017] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The purposes and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0018] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 A flowchart of a method for predicting a fault in an energy storage liquid cooling system provided by an embodiment of the present invention;

[0021] Figure 2 A flowchart of another method for predicting faults in an energy storage liquid cooling system provided by an embodiment of the present invention;

[0022] Figure 3 A schematic diagram of a comparison result corresponding to an adaptive segmentation provided by an embodiment of the present invention;

[0023] Figure 4 A schematic diagram of the effect corresponding to a dynamic convolution kernel provided by an embodiment of the present invention;

[0024] Figure 5 A schematic diagram of the effect of online incremental learning provided by an embodiment of the present invention;

[0025] Figure 6 A schematic structural diagram of a fault prediction device for an energy storage liquid cooling system provided by an embodiment of the present invention;

[0026] Figure 7 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0028] The embodiments of the present invention provide a method and device for predicting faults in an energy storage liquid cooling system to solve the above-mentioned problems.

[0029] To facilitate understanding of this embodiment, the energy storage liquid cooling system fault prediction method disclosed in the embodiment of the present invention is first introduced in detail. Figure 1 The flowchart corresponding to the embodiment of the present invention is shown. Figure 1 , the method comprises the following steps:

[0030] Step S102: Obtain operation monitoring parameters corresponding to the preset energy storage liquid cooling system.

[0031] Operational monitoring parameters include those corresponding to pre-set key equipment in the energy storage liquid cooling system. These parameters are primarily collected in real time via multiple types of sensors deployed on key components, such as coolant lines, battery modules, and circulation pumps. These sensors include temperature, pressure, flow, and voltage / current sensors. Data is uploaded to a local server or cloud database via an industrial bus (such as the CAN bus or Modbus protocol) or a wireless transmission module (such as LoRa or 5G). The sampling frequency is dynamically adjusted based on the equipment's operating status, ranging from 1Hz under normal operating conditions to 10Hz under abnormal conditions to ensure fault signature capture. Raw data is stored in a time series format, containing fields such as device ID, timestamp, sensor type, and value, forming the initial monitoring dataset.

[0032] Furthermore, during the data acquisition phase, the raw signal can be initially filtered and outliers removed. For example, a sliding mean filter can be used to eliminate transient interference noise, and threshold detection can be used to filter out obvious abnormal data, reducing interference in subsequent processing. Furthermore, data storage can adopt a time-based archiving strategy, dividing files by hour or day to facilitate subsequent segmented processing and historical data backtracking.

[0033] Step S104 : Based on the local noise level and time scale variation of the operation monitoring parameter, data of the operation monitoring parameter is divided to obtain segmented data corresponding to a plurality of time periods.

[0034] Step S106: Determine the denoising strength corresponding to the segmented data according to the noise intensity corresponding to the segmented data.

[0035] Step S108 , performing denoising processing on the segmented data based on the denoising strength, and determining the operation monitoring parameters based on the parameters to be measured corresponding to each time scale.

[0036] The system exhibits different dynamic behaviors during different operating phases (startup, stable operation, load changes, and pre-fault warnings). Fault types also vary in their timescales. For example, sudden faults (such as pipeline ruptures) have short timescales and high noise levels, while slow-onset faults (such as decreased cooling efficiency) have long timescales and weak signatures. Furthermore, the data collected by sensors over different time periods is affected by factors such as environmental interference, equipment status, and sampling frequency, resulting in significant fluctuations in noise levels. Furthermore, under normal operating conditions, early fault signals are very weak and are often overwhelmed by high-frequency noise or background fluctuations.

[0037] In order to solve the problem of insufficient energy for fault feature extraction in the prior art, an embodiment of the present invention segments the data based on the local noise differences in the data, so that the noise intensity, trend change, and time scale within the same segment are relatively consistent. Based on this adaptive data partitioning mechanism, subsequent processing can be processed separately according to "quasi-homogeneous" data blocks, thereby improving the accuracy of denoising and feature extraction. Furthermore, the denoising strength is automatically adjusted according to the noise intensity corresponding to each segment of data, and personalized denoising operations are performed according to its specific denoising strength, so as to effectively suppress noise and avoid the loss of useful fault features, ensuring that features at each time scale can be accurately extracted. In summary, the final output data is clean test parameters at multiple time scales, which can provide high-quality input data for the neural network, improving the interpretability of features and the generalization ability of the model.

[0038] In step S110 , the parameters to be measured are input into a pre-trained neural network model, and the neural network model is used to perform classification prediction on the parameters to be measured, so as to determine the fault category prediction result corresponding to the parameters to be measured.

[0039] Step S112: Determine the fault state corresponding to the energy storage liquid cooling system according to the fault category prediction result.

[0040] Furthermore, the processed data is input into a preset neural network model (such as an LSTM, GRU, Transformer, or CNN-LSTM hybrid model) for inference. The model can output a predicted category, which can be different fault levels or types. For example, the fault category (e.g., coolant leak, pump blockage), severity level, and location. The predicted category can be determined based on the sample labels corresponding to the training sample set. Data labels can be determined by associating sensor data with fault events through timestamp alignment based on equipment operation logs and fault maintenance records. For example, when the system records a fault alarm (e.g., overtemperature protection triggering) or a manual maintenance report, the monitoring data is backtracked and the data labels corresponding to the time window of the fault occurrence (e.g., 30 minutes before the fault to 5 minutes after the fault) are annotated. In one embodiment, the annotated data can be divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning, and the test set is used for final performance evaluation to ensure model generalization.

[0041] Furthermore, the model can output a probability distribution for fault categories. When the highest probability exceeds a set threshold (e.g., 0.9), a graded alarm is triggered, and the fault location and recommended remediation measures are simultaneously pushed to the operation and maintenance platform. Historical prediction results and device status data can also be stored in a knowledge base, allowing for periodic incremental updates to the model.

[0042] In summary, the embodiment of the present invention provides a method for predicting faults in an energy storage liquid cooling system, which realizes adaptive segmentation processing of the original data by dynamically analyzing the local noise level and time scale changes of the monitoring data. It can better capture the characteristics of different fault modes, especially faults with significant time scale differences (such as sudden faults and slow-changing faults). In addition, the embodiment of the present invention also dynamically adjusts the denoising strategy according to the characteristics of each segmented data, which can not only effectively remove noise but also retain useful information to the maximum extent, which helps to improve the data quality of subsequent model training. The embodiment of the present invention ensures that even in a high-noise environment, weak but critical fault features can be accurately extracted by precisely controlling the denoising process, which is crucial for improving the sensitivity and specificity of the fault diagnosis system.

[0043] Furthermore, based on the above embodiment, the embodiment of the present invention also provides another method for predicting failure of an energy storage liquid cooling system. Figure 2 The flowchart corresponding to the embodiment of the present invention is shown. Figure 2 , the method comprises the following steps:

[0044] Step S202: Obtain operation monitoring parameters corresponding to the preset energy storage liquid cooling system.

[0045] Step S204 : Based on the local noise level and time scale variation of the operation monitoring parameter, data of the operation monitoring parameter is divided to obtain segmented data corresponding to multiple time scales.

[0046] Operational monitoring data often exhibits non-stationary characteristics over time. For example, equipment may exhibit different performance during different operating phases (such as startup, stable operation, and shutdown). Furthermore, changes in external environmental conditions (such as temperature and humidity) can also contribute to data non-stationarity. Noise levels can vary significantly across time periods. Some periods may experience high background noise, while others are relatively clear. Equipment failures or other emergencies can cause sudden increases in local noise, impacting the accuracy of global analysis. To address this data characteristic, embodiments of the present invention propose an adaptive wavelet denoising and dynamic time segmentation strategy. This strategy calculates local noise characteristics and uses an adaptive segmentation method to segment the data. Furthermore, the denoising threshold can be dynamically adjusted based on changes in the data's timescale (such as short-term fluctuations and long-term trends). This allows for flexible adaptation to noise intensity and timescale variations under different fault modes, avoiding the potential information loss associated with fixed-threshold denoising. This improves fault feature retention and diagnostic accuracy.

[0047] In a specific implementation, the operation monitoring parameters are first segmented through the following steps, ensuring that the data within each segment has similar noise levels and time scales, adapting to the characteristic distributions of different fault modes. Specifically, the data noise difference within the current time period is determined based on the data noise level corresponding to each time point in the preset time period of the operation monitoring parameter. Based on the data noise difference, the time segment points corresponding to the operation monitoring parameter are determined. Based on the time segment points and the preset time segment length penalty term, the operation monitoring parameter data is segmented to obtain segmented data corresponding to multiple time scales.

[0048] In one embodiment, time periods can be dynamically divided by K-means clustering. Combining the above steps, the clustering objective function can be expressed as:

[0049]

[0050] Where, represents the minimization objective function; Time period T k time point within; For the original Real-time monitoring data of energy storage liquid cooling system; To control the segment length penalty term, for example, it is set to 0.3. Among them, the embodiment of the present invention can use a clustering algorithm to cluster the data to obtain the cluster to which each time point belongs. Based on the clustering results, the positions where adjacent time points belong to different clusters are potential time segmentation points. A length penalty term can be introduced in the clustering process to prevent segmentation from being too frequent. Furthermore, data is segmented based on the length constraints of the time segmentation points and the length penalty term to obtain multiple segmented data. Since the segmented data is segmented according to the noise difference corresponding to the time point, each segmented data has the same time scale, such as short-term time scale segmented data corresponding to transient fault mode; medium-term time scale segmented data corresponding to sudden fault; long-term time scale segmented data corresponding to long-term trends and structural changes. Compared with the traditional fixed threshold denoising processing under the traditional sliding window segmentation, the embodiment of the present invention is based on adaptive segmentation and applies non-fixed threshold wavelet denoising processing, which can adapt to different fault modes to better capture details and improve the retention of fault features and diagnostic accuracy.

[0051] It should be noted that the time scales of the failure modes of the energy storage liquid cooling system vary significantly. Setting values ​​of different sizes can prevent the instability of noise estimation caused by too short segments or the masking of time-varying features caused by too long segments, thereby adapting to non-stationary time series characteristics.

[0052] Step S206 : determining the denoising strength corresponding to the segmented data according to the noise intensity corresponding to the segmented data.

[0053] The noise intensity in the monitoring data of the energy storage liquid cooling system varies depending on the fault mode. For example, the noise distribution for sudden temperature change and gradual pressure change faults is different. This embodiment of the present invention dynamically adjusts the denoising strength based on the noise intensity of the local data segment (i.e., the aforementioned segmented data). The calculation method is expressed as:

[0054]

[0055] Among them, the embodiment of the present invention combines the range and variance of the data within the segment, reduces the suppression coefficient when the noise intensity is high (large variance), and retains more potential fault features; enhances the suppression when the noise intensity is low (small variance) to avoid over-smoothing, thereby solving the problem of excessive suppression of weak fault features by fixed threshold wavelet denoising.

[0056] In the above formula, is the dynamic noise suppression coefficient, T k is the kth time period; Indicates the The maximum value of the monitoring data within the time period; Indicates the minimum value of the monitoring data in the kth time period; is the variance of the data in the kth time period, which represents the volatility of the data in this period.

[0057] Step S208 , performing denoising processing on the segmented data based on the denoising strength, and determining the operation monitoring parameters based on the parameters to be measured corresponding to each time scale.

[0058] Among them, the embodiment of the present invention pre-adaptively segments the data according to the noise difference of the data, so that each segmented data has a similar noise level and time scale. Furthermore, each segmented data is denoised according to the denoising strength corresponding to the different segmented data, thereby obtaining the measured parameters after the denoising of the operation monitoring parameters. Among them, the denoising of the data by the embodiment of the present invention can not only distinguish different fault modes, but also effectively reduce the loss of effective information corresponding to each fault mode. Specifically, the step of denoising the segmented data based on the denoising strength refers to the following formula:

[0059]

[0060] Where, is the denoised monitoring data of the energy storage liquid cooling system at time t; For the original The monitoring data of the energy storage liquid cooling system at all times. K is the number of segments, and the number of segments is dynamically adjusted to adapt to the time scale differences of different failure modes. is the Daubechies wavelet basis function. In the embodiment of the present invention, the wavelet odd function is used to extract the characteristics of the local signal of the segmented data to denoise the data. This method can adapt to non-stationary signals. As the center, is the window width, extracting the local signal. t is the time variable; is the center point of the kth segment; is the kth window width; is the mean of the kth segment of data, which is used to measure the center position of each segment of data.

[0061] To verify the effectiveness of this technology, the embodiment of the present invention also uses a dual-axis time series diagram to compare the performance of different time segmentation strategies. Taking the temperature data of the energy storage liquid cooling system as an example, Figure 3A schematic diagram shows the comparison results of the adaptive segmentation provided by an embodiment of the present invention. State data for simulated temperature signals undergoing sudden and gradual changes is used to verify the effectiveness of the segmentation strategy. The signal amplitude on the vertical axis is the absolute temperature (in degrees Celsius), representing the temperature increase relative to a reference temperature, which is the ambient temperature. The main figure shows a composite fault signal consisting of high-frequency oscillations and a gradual trend, superimposed with impulse noise. The gray background curve represents noisy observation data, while the black dashed line represents the ideal true signal. The vertical dividing lines are colored to distinguish the segmentation results of different methods. Experimental results show that the green dashed line represents the adaptive segmentation of the proposed technique, which accurately segments near the fault mode transition point (the starting point of the signal sudden change). The red solid line represents equally spaced fixed segments, which exhibit significant offset at the onset of a gradual fault. The orange dashed-dotted line represents the sliding window segmentation, which captures sudden changes but produces a large number of redundant segments. The zoomed-in view highlights the starting area of ​​a gradual fault, showing that the segmentation lines of the proposed technique accurately align with the signal trend transition point, while the comparison methods lag or advance, demonstrating that the dynamic clustering strategy can adaptively match the time-varying characteristics of the fault mode. It should be noted that the monitoring data of the energy storage liquid cooling system also includes many other types of parameters, such as pressure data, flow data, liquid level data, etc. Different types of monitoring data are Figure 3 The signal amplitude units in the data are different. When the monitoring data is pressure data, the unit can be Pa (Pascal), kPa (Kilopascal), or MPa (Megapascal); when the monitoring data is flow data, the unit can be L / min (liters / minute) or m³ / h (cubic meters / hour). Furthermore, other types of monitoring data can be expressed in corresponding international standard units, which are not explained here.

[0062] In step S210, the parameters to be measured are input into a pre-trained neural network model, and the time series feature dynamic fusion layer of the neural network model is used to capture the features of the parameters to be measured using multiple convolution kernels of different scales to determine the local target features corresponding to the parameters to be measured.

[0063] Embodiments of the present invention use a neural network model for data classification and prediction. The neural network employed in these embodiments utilizes a layered temporal feature extraction architecture, comprising an input layer, a dynamic temporal feature fusion layer, an adaptive regression unit, and a multi-scale residual connection module. In one embodiment, the input layer receives denoised and segmented temporal data of dimension [time step × number of sensor channels]. The dynamic temporal feature fusion layer consists of parallel dynamic convolution kernels and an attention mechanism, extracting local features using three sets of convolution kernels of different scales and weightedly fusing them with temporal attention. The adaptive regression unit comprises four expert subnetworks, each consisting of three fully connected layers with a gating mechanism dynamically selecting activation paths. The multi-scale residual connection module spans each hidden layer, fusing the pooled features of the first three layers to enhance gradient propagation. The output layer uses a softmax function to generate fault category probabilities, with the category with the highest fault probability being the final classification category. It should be noted that this neural network architecture can be flexibly expanded, for example, by increasing the number of fully connected neural network layers, adding ordinary convolution layers, or adding pooling operations, all of which fall within the scope of this technology.

[0064] In the specific implementation, first, after the input layer receives the data, the dynamic fusion layer of the time series features determines the local target features corresponding to the parameters to be measured. Since the monitoring data of the energy storage liquid cooling system has multivariable coupling and non-stationary time series characteristics, conventional technologies directly stack time series data or extract statistical features, which cannot capture the time-varying coupling relationship of fault characteristics. The present invention constructs a time series feature extraction layer that integrates the dynamic convolution kernel and the attention mechanism. By constructing a time series feature extraction layer that combines the dynamic convolution kernel and the attention mechanism, the time-varying coupling relationship of fault characteristics can be effectively captured. In this process, the dynamic convolution kernel weights the features at each moment through the time attention mechanism, thereby improving the characterization ability of fault characteristics.

[0065] The specific implementation includes the following steps:

[0066] 1) The characteristic parameters of the measured parameters at each moment are nonlinearly mapped by a preset multi-layer perceptron to determine the query vector corresponding to the measured parameters at each moment.

[0067] 2) According to the attention mechanism of the query vector and the key vector of the preset dynamic convolution kernel, the convolution kernel attention weight corresponding to the feature parameter of the parameter to be measured at each moment is determined.

[0068] It should be noted that there is a time-varying coupling relationship between multiple variables (such as temperature, pressure, and flow) in the energy storage liquid cooling system. For example, a sudden change in the coolant flow rate causes temperature fluctuations. generated by a multi-layer perceptron nonlinear mapping) and the key vector The attention mechanism dynamically adjusts the weight of the convolution kernel on the features at different time points, thereby capturing the instantaneous coupling relationship of the fault features. is the query vector at time t, and is obtained as follows: ; It is a multi-layer perceptron, such as a 3-layer perceptron network.

[0069] Furthermore, define is the i-th dynamic convolution kernel at time t, and the convolution kernel weight is dynamically generated according to the current moment characteristics. The calculation method is expressed as .

[0070] in, is the Softmax function, which is used to calculate the weight of the convolution kernel to help the model identify which moments are more important; K i is the i-th key vector, representing the key vector of the dynamic convolution kernel, which is a training parameter; K i The transpose of The dimension of the query vector and key vector is set to 64 by default, which controls the computational complexity of the attention weight. is the i-th value vector, representing the adjustment parameters of the convolution kernel, which is the training parameter.

[0071] 3) According to the convolution kernel weight and the temporal attention weight corresponding to the feature parameter, the temporal characteristics of the parameter to be measured are captured and the local target features corresponding to the parameter to be measured are determined.

[0072] Determine the local target features according to the following formula:

[0073]

[0074] Where, is the fusion feature at the tth moment, which is represented by the result of dynamic fusion of temporal features. is the number of dynamic convolution kernels, indicating that multiple convolution kernels are used in the model to capture different types of temporal features, for example, it is set to 3; Represents a convolution operation. is the i-th dynamic convolution kernel at the t-th moment.

[0075] in, is the time attention weight at the tth moment, and the features in the local time window are weighted and fused. The calculation method is expressed as:

[0076]

[0077] is the local time index within the time sliding window; is the local time window length; Indicates the The convolution kernel is The attention coefficient of the moment; Represents the first Monitoring data at the moment. It should be noted that the fault characteristics may be locally sudden, such as the sudden drop in flow at the moment the cooling pump stops. The embodiment of the present invention is based on the use of a time sliding window to extract the local context information of the measured parameters to determine the time attention weight. Combined with the above formula, the embodiment of the present invention extracts the local context through a sliding window, combines the Sigmoid function to adaptively assign weights, strengthens the feature contribution of key time points, and thus enhances the sensitivity to sudden faults while retaining the long-term trend of slow-changing faults; is the Sigmoid activation function.

[0078] Furthermore, the embodiment of the present invention also uses a time-frequency domain joint analysis method to analyze the analytical capabilities of different methods for non-stationary signals. Figure 4 Schematic diagram showing the effect of the dynamic convolution kernel corresponding to the embodiment of the present invention. Figure 4 The experiment focuses on verifying the effect of dynamic convolution kernel and attention mechanism on capturing time-varying features, through time domain waveform, power spectrum density and time-frequency spectrum. Figure 3 Re-verify. Figure 4 In the figure, the time domain curve above shows that the blue signal of this technology eliminates the low-frequency fluctuation noise in the gray original signal while retaining the attenuated oscillation characteristics of the original signal; the spectrum diagram on the left below shows that the results of this technology maintain clear spectral peaks at 2Hz and 5Hz. Among them, the horizontal axis represents frequency, and the vertical axis represents power spectral density (PSD for short), with the unit of decibel per hertz (dB / Hz), which represents the decibel power value within the bandwidth of each hertz. However, the spectrum line of the traditional method shows frequency diffusion phenomenon. For the time-frequency diagram on the right below, the embodiment of the present invention uses color intensity to represent the signal energy distribution. The spectrum corresponding to this technology completely retains the short-time pulse characteristics of the high-frequency transient component in the time dimension, proving the ability of the dynamic convolution kernel to adjust the weight according to local characteristics, and overcoming the problem of time-frequency resolution loss caused by the fixed convolution kernel in the traditional method.

[0079] In step S212 , the adaptive regression unit of the neural network model uses a preset expert selection matrix to perform adaptive activation path selection on the parameters to be measured, and determines a high-dimensional feature representation corresponding to the parameters to be measured.

[0080] In the fault diagnosis task of the energy storage liquid cooling system, the time series regression characteristics of different fault modes vary significantly. For example, the time series regression characteristics of different fault modes may show slow changes and sudden changes. The traditional fully connected layer cannot adapt to the differences in regression characteristics of different fault modes, which easily leads to underfitting of rare fault modes. The present invention constructs an adaptive regression unit that can adaptively select appropriate regression parameters according to different fault states. Through the expert selection mechanism, the model can automatically adjust the parameters in the regression process under various fault modes, thereby adapting to different types of fault characteristics and improving the generalization ability of fault diagnosis.

[0081] The specific implementation includes the following steps:

[0082] 1) Using the preset expert selection matrix, the parameters to be tested are mapped into expert probability distribution.

[0083] 2) Based on the expert probability distribution, determine the expert units corresponding to the characteristic parameters of the parameter to be measured at each moment.

[0084] Among them, the regression characteristics of different fault modes are significantly different, such as the difference between the slow change situation and the sudden change situation. Mapping to expert probability distribution, so that the model automatically switches to the regression parameters corresponding to the failure mode and , thereby solving the underfitting problem of the traditional fully connected layer for rare fault modes and improving the model's diagnostic ability for long-tail categories (low-frequency faults). Furthermore, the most suitable expert unit is dynamically selected based on the input features. Each element of the expert selection matrix represents the degree of influence or correlation of a specific feature on a specific fault mode. The dimension of the expert selection matrix in the embodiment of the present invention is , is the total number of experts, representing the number of possible failure modes. Each row corresponds to a different feature, each column to a different failure mode, and each element reflects the relevance or contribution of the i-th feature to the j-th failure mode. These values ​​can be determined in a variety of ways, including but not limited to expert opinion, experimental data, and historical failure data.

[0085] In the embodiment of the present invention, the expert unit can be determined by the following formula:

[0086]

[0087] Among them, s is the expert unit index of the expert selection mechanism, Indicates the index value corresponding to the maximum probability. is the Softmax function. represents the probability distribution of expert selection, where each element corresponds to the probability of selecting the sth expert. E is the expert selection matrix, and is the training parameter, representing the linear transformation matrix from feature space to expert selection.

[0088] 3) Based on the regression parameters corresponding to the expert units, a nonlinear transformation is performed on the parameters to be measured to determine the high-dimensional feature representation corresponding to the parameters to be measured.

[0089] In the specific implementation, refer to the following formula:

[0090]

[0091] Where F(t) is the output feature of the dynamic fusion layer of temporal features. For the Tier The output of each expert unit represents the nonlinear transformation result for a specific fault mode; is the output feature of the l-1 layer; is the ReLU activation function; is the weight matrix of the s-th expert unit in the l-th layer, and is the training parameter; is the bias term of the s-th expert unit in the l-th layer, and is the training parameter. s is the expert unit index of the expert selection mechanism, which is determined by the above steps.

[0092] It should be noted that the role of the adaptive regression unit is to make the parameters and Adaptive adjustment during model training automatically adjusts the regression process according to different failure modes to avoid underfitting of rare failure modes.

[0093] In step S214, based on the preset stride convolution operation, the time scale features corresponding to the local target features and the high-dimensional feature representations are displayed and extracted.

[0094] It should be noted that the failure of the energy storage liquid cooling system may involve multi-time scale dynamic characteristics, such as long-term cooling efficiency decline and short-term flow fluctuation. The embodiment of the present invention also presets the step size, such as The convolution operation explicitly extracts short-term (m=1), medium-term (m=2), and long-term (m=3) features, thereby enhancing the model's ability to model complex failure modes. In one embodiment, define Indicates the step size is One-dimensional convolution.

[0095] Step S216, based on the contribution of the parameter to be measured to the dynamic fusion layer of the temporal feature or the adaptive regression unit, the time scale feature corresponding to the local target feature and the time scale feature corresponding to the high-dimensional feature representation are adaptively fused to determine the target fusion feature corresponding to the parameter to be measured.

[0096] Since the monitoring data of the energy storage liquid cooling system has dynamic characteristics and long-term dependencies at multiple time scales, the gradient vanishing problem is serious in deep neural network training. Conventional residual connections ignore the differences in the contribution of features at different time scales. To solve this problem, the present invention adopts hierarchical multi-scale residuals to combine multi-layer features to effectively solve the common gradient vanishing problem in deep neural networks. Among them, the embodiment of the present invention uses residual terms of different scales, so that the network can learn features at different time scales, thereby improving the robustness and generalization ability of the model.

[0097] Specifically, the residual connection is performed on the above time scale features based on the residual term to determine the corresponding target fusion features. The process is expressed as:

[0098]

[0099] Where, For the Layer residual term, enhancing the cross-layer propagation of gradient. It is the maximum time scale number of the residual connection. For example, it is set to 3, which means the features of the first 3 layers are fused. is the output feature of the lmth layer. Indicates a step size of 2 m One-dimensional convolution. is the residual weight of the m-th scale in the l-th layer.

[0100] In the specific implementation, the residual weight is dynamically adjusted according to the correlation between the current layer and the historical layer features, and the calculation method is expressed as: , which measures the contribution of the mth-scale feature to the current layer. This can be adaptively weighted using a gating mechanism. It's important to note that features at different time scales in deep networks contribute differently to fault diagnosis. For example, long-term dependencies are more important for temperature control failure. Using a sigmoid function as a gating mechanism allows for adaptive weight allocation, preventing the dilution of important features caused by fixed residual connections, thereby alleviating the vanishing gradient problem while preserving the discriminative nature of multi-scale fault features.

[0101] in, is the weight matrix of the mth scale, and the concatenated features Mapped to scalar weights , controls the fusion ratio of features of different scales. Among them, The step size is 2 m Average pooling operation, compressing the time dimension to 1 / 2 m , extracting long-term dependency features. The representation concatenates the pooled features with the current layer features to provide multi-scale context information. is the output feature of the lth layer.

[0102] Step S218: perform fault category probability prediction on the target fusion feature to determine the fault category prediction result corresponding to the parameter to be measured.

[0103] In combination with the above steps, the embodiment of the present invention uses the Softmax function in the output layer of the neural network model to generate the fault category probability, and takes the category with the highest fault category probability as the final classification category.

[0104] Step S220: Determine the fault state corresponding to the energy storage liquid cooling system according to the fault category prediction result.

[0105] In summary, the present invention proposes an adaptive wavelet denoising and dynamic time segmentation strategy. By calculating local noise characteristics and time-scale variations, the denoising threshold is dynamically adjusted, and an adaptive segmentation method is used to segment the data. Compared with traditional fixed-threshold wavelet denoising and sliding window segmentation, this method can flexibly adapt to the differences in noise intensity and time scale under different fault modes, avoiding the information loss that can occur with fixed-threshold denoising, thereby improving the preservation of fault characteristics and diagnostic accuracy.

[0106] Furthermore, in order to cope with the multivariate coupling and non-stationary characteristics of the monitoring data of the energy storage liquid cooling system, the present invention adopts a dynamic fusion layer of time series features, which combines the dynamic convolution kernel with the temporal attention mechanism to effectively capture time-varying fault characteristics. The adaptive regression unit adopts an expert selection mechanism, which can adaptively adjust the parameters in the regression process according to different fault modes, thereby improving the model's ability to recognize various types of fault modes.

[0107] Furthermore, the embodiment of the present invention also designs a training method for a neural network model. The training steps of the embodiment of the present invention are as follows:

[0108] 1) Obtain a pre-built training sample set.

[0109] 2) The training sample sets are input into the preset neural network model in batches, and the classification error, regression error and total loss corresponding to the expert diversity constraint term of the training sample sets are calculated through the preset hybrid loss function.

[0110] Conventional single cross entropy loss cannot handle the problems of fault category imbalance and regression error accumulation; the present invention adopts a hybrid loss function to solve the problems of fault category imbalance and regression error accumulation by combining classification loss, regression error and expert unit parameter constraints, thereby improving the accuracy of fault diagnosis.

[0111] The specific implementation includes the following steps:

[0112] a-Determine the weighted regression loss weight based on the number of samples corresponding to different sample categories in the preset training sample set.

[0113] Among them, the sample sizes of common and rare categories in the energy storage liquid cooling system fault data are extremely unbalanced. In one embodiment, the regression loss of high-frequency faults can be downweighted by an exponential decay function to avoid overfitting of the majority class, thereby alleviating the model bias caused by category imbalance and improving the diagnostic accuracy of rare faults.

[0114] In one embodiment, the weighted regression loss can be calculated by the following formula:

[0115]

[0116] Referring to the above formula, the embodiment of the present invention dynamically adjusts the weight of the regression loss according to the number of samples to alleviate the problem of class imbalance. Characterization by number of class samples Adjust the regression loss weight, is an exponential function with a natural constant as its base; is the category balance factor, for example, set to 0.1. is the total number of fault categories. For the Class prediction value; For the Class true value; is the L2 norm. is the number of samples in the kth category. is the regression loss weight coefficient, for example, set to 0.3.

[0117] b- According to the preset parameter vector constraint algorithm, parameter space differentiation constraints are performed on the parameter vectors of the preset multiple expert units respectively to determine the expert diversity constraint loss corresponding to the training sample set.

[0118] The expert unit is used to learn the preset target fault mode features in the training sample set based on the corresponding parameter space. The embodiment of the present invention also designs an expert diversity constraint loss to encourage different expert units to learn differentiated features and prevent multiple expert units from learning similar fault mode features. For example, different experts all focus on temperature features and ignore pressure features. In one embodiment, the cosine similarity between parameter vectors can be minimized to encourage expert units to focus on differentiated features, thereby improving the model's ability to analyze complex faults. Specifically, it is characterized as: .

[0119] in, is the expert diversity weight coefficient, which is used to constrain the orthogonality of parameters of different expert units. For example, it is set to 0.2. For the sth and The angle between the parameter vectors of the expert units, The representation constrains the differences in parameter spaces of different expert units and avoids expert mode collapse by minimizing the cosine similarity between expert parameters.

[0120] c-Calculate the hybrid loss function corresponding to the training sample set based on the expert diversity constraint loss, weighted regression loss weight, and preset classification loss.

[0121] In one embodiment, cross entropy loss can be used as the classification loss. In combination with the above steps, the calculation method of the hybrid loss function can refer to the following formula:

[0122]

[0123] Where, is a mixed loss function that represents the overall training objective. is the classification loss; is the classification loss weight coefficient, for example, set to 0.2. is the regression loss weight coefficient, for example, set to 0.3; is the expert diversity weight coefficient, which constrains the orthogonality of parameters of different expert units, for example, it is set to 0.2.

[0124] 3) Update the parameters of the neural network model based on the total loss.

[0125] The parameter update corresponding to the embodiment of the present invention refers to the following formula:

[0126]

[0127] Where, is the updated model parameter, representing the parameter state after incremental learning. is the current model parameter, representing the parameter state before the update. is the learning rate, for example, set to 0.001. is the gradient of the hybrid loss function with respect to the model parameters. is the elastic weight coefficient, for example, set it to 0.3; is a historical model parameter that characterizes the parameter state before the last update and is used for elastic weight constraints. Referring to the above formula, the embodiment of the present invention obtains the model parameters corresponding to the neural network model at each iteration and constrains the parameter update amplitude of the neural network model based on the parameter state corresponding to the model parameters. Furthermore, the total loss corresponding to the pre-stored memory samples is extracted from a pre-set memory library, and the parameters of the neural network model are updated based on the total loss corresponding to the memory samples. It should be noted that the memory samples are samples in the training sample set.

[0128] It should be noted that The role of is to limit the parameter update range and retain historical knowledge. Since the data of new fault modes may destroy the existing fault diagnosis capabilities, the elastic weight coefficient Constraining the parameter update direction prevents catastrophic forgetting, thereby enabling the model to quickly adapt to new fault modes while maintaining stable diagnostic performance for existing faults.

[0129] Furthermore, the above formula also includes Represents the gradient of the hybrid loss function calculated based on the i-th sample stored in the memory bank with respect to the model parameters. A memory bank that stores representative samples ; is the i-th sample stored in the memory bank. is the memory bank gradient weight coefficient, which represents the contribution strength of the memory sample to the parameter update, for example, it is set to 0.5.

[0130] Furthermore, since the failure modes of the energy storage liquid cooling system may evolve over time, such as equipment aging causing fault feature shifts, existing technologies usually require full retraining of the model when facing new failure modes, resulting in a waste of computing resources and time delays. The embodiment of the present invention also adopts an incremental learning method based on dual memory banks, which can eliminate the need for full retraining when new failure modes occur, reducing computing and time overhead. Through dual memory bank updates, the model can continuously adapt to new failure modes and maintain stable performance.

[0131] In a specific implementation, the embodiment of the present invention determines the new sample corresponding to the training sample set through online learning. Further, the new sample can be measured by the kernel function and memory bank The similarity between the two sets of data is calculated, and only the difference samples are retained to ensure that the memory library covers the diversity of data distribution, avoid memory library sample redundancy during incremental learning, and improve the model's generalization ability for edge cases. Furthermore, based on the corresponding difference samples, the memory library is updated. The memory library update strategy refers to the following formula:

[0132]

[0133] Where, For the assignment operation; is the new sample at the current moment, representing the new fault data received during the online learning process; The similarity threshold is set to 0.8, for example, to ensure that the memory library covers samples at the edge of the data distribution; is a kernel function based on feature space similarity, which represents the similarity measure between samples. The calculation method is expressed as . is the kernel bandwidth of the kernel function based on feature space similarity, which controls the similarity decay rate, for example, set it to 0.1.

[0134] 4) During the parameter update process of the neural network model, the feature elements of the training sample set are randomly discarded based on the preset time-sensitive discarding strategy.

[0135] In order to make the network structure slightly different at each iteration and force the model to learn more robust features, the existing technology achieves this by randomly discarding some neurons. However, traditional Dropout can easily destroy the temporal continuity in the monitoring data of the energy storage liquid cooling system, which can easily lead to the model being unable to effectively capture the timing dependency. The present invention adopts a time-sensitive Dropout based on feature importance. While maintaining temporal continuity, it suppresses overfitting by dynamically adjusting the probability of Dropout and improves the generalization ability of the model. In specific implementation, the probability of random discarding in the embodiment of the present invention is calculated based on the feature amplitude and diversity corresponding to the feature elements. Specifically, refer to the following formula:

[0136]

[0137] Where, It is a time-sensitive Dropout function that retains residual information when features are discarded to maintain time continuity. is the multi-scale residual term at the t-th moment of the l-th layer, which is calculated by the multi-scale residual connection. is the feature vector of the lth layer at the tth moment; is element-wise multiplication; is the mask vector, representing the probability Randomly drop features.

[0138] Among them, the calculation method of the mask vector is expressed as:

[0139]

[0140] in, Indicates that it obeys a certain distribution; is the Bernoulli distribution, Characterize each feature element with probability This distribution can be seen as a transition from the Bernoulli distribution (B(1, )) where the probability of success (i.e. the neuron is retained) is (1- ), the probability of failure (i.e., the neuron is discarded) is ( ). It can ensure that the network does not rely too much on any single neuron, promoting more effective feature learning.

[0141] Among them, the above is the feature retention probability, and the discard probability is dynamically adjusted according to the feature amplitude and diversity. The calculation method is expressed as , characterizing the feature retention probability to be negatively correlated with amplitude and diversity, suppressing overfitting of high activation or single features. is the characteristic entropy, and the calculation method is expressed as , characterizing the diversity of feature distribution. is the first feature of the lth layer at the tth moment dimension; It is a logarithmic function, and the default base is 10; is the dimension of the feature vector at the tth moment in the lth layer. is the first control strength, representing the control strength of the characteristic amplitude; is the second control strength, which characterizes the control strength of feature diversity.

[0142] It should be noted that high activation features (such as a sudden increase in pressure value) in the monitoring data of the energy storage liquid cooling system may contain key fault information, but a single feature (such as only temperature anomaly) is prone to overfitting. and entropy term , suppressing the overfitting tendency of high-activation or low-diversity features, and improving the model's robustness to noise interference while maintaining temporal continuity.

[0143] 4) Until the neural network model meets the preset training requirements, the current optimal weight parameters are selected to generate the target neural network model.

[0144] In combination with the above steps, a small batch gradient descent method can be used to update the parameters with the Adam optimizer. Among them, the initial learning rate can be set to 0.001 and dynamically adjusted using the cosine annealing strategy. In one embodiment, 256 time window samples can be input per batch, and the total loss of classification error, regression error and expert diversity constraint terms is calculated by a mixed loss function, and the network weights are updated by backpropagation. During the training process, the accuracy and F1 score are evaluated on the validation set after each iteration. If the validation loss does not decrease for 10 consecutive rounds, the training is terminated early to prevent overfitting. The final model selects the weight parameters with the best performance on the validation set, and freezes the network for online deployment.

[0145] In summary, the embodiments of the present invention solve the problem of bias of classification loss on unbalanced data in traditional methods by adopting a hybrid loss function, especially for the processing of rare fault categories. By optimizing the classification and regression errors, the accuracy of fault diagnosis is improved, and through the expert diversity constraint, the generalization ability of the model under multiple fault modes is enhanced.

[0146] In addition, the traditional Dropout method may destroy the temporal nature of the data when processing time series data. The present invention proposes a time-sensitive Dropout strategy that can maintain time series continuity while effectively suppressing overfitting. In addition, the probability of Dropout is adjusted in combination with feature importance, making the model more robust and able to maintain good performance in the presence of noise interference and missing data. The present invention also adopts an incremental learning method based on dual memory banks. When a new fault mode appears, there is no need for full retraining. The new fault mode is adapted through incremental updates, which solves the problem that traditional models need to be retrained when facing new modes. It significantly reduces computational and time overhead, while avoiding catastrophic forgetting and maintaining the long-term stability and accuracy of the model.

[0147] Furthermore, the embodiment of the present invention also uses a line chart with confidence intervals to evaluate the stability of the online incremental learning strategy. Figure 5 The following figure shows the effect of online incremental learning corresponding to the embodiment of the present invention. In particular, the embodiment of the present invention compares the traditional full update method with the dual memory bank strategy of the present technology based on the scenario where new failure modes continue to appear during the continuous operation of the equipment. Figure 5 The horizontal axis represents the number of online learning rounds, and the vertical axis, diagnostic accuracy, reflects the model's comprehensive diagnostic capability for both new and old fault modes. The green curve corresponding to this technology has a confidence interval of 100% in all three sets of experiments, which is higher than that of the comparison method and higher than the threshold confidence interval (95% confidence interval) commonly used in this field to assess whether the confidence level has reached an excellent level. This shows that the sample diversity maintenance mechanism of the memory bank effectively balances the absorption of new knowledge and the retention of old knowledge. The red curve of the traditional method shows a significant decline after the fifth round, and the confidence interval continues to expand, indicating performance fluctuations caused by catastrophic forgetting, indicating that the elastic weight constraint of this technology improves system stability.

[0148] Furthermore, based on the above embodiment, the embodiment of the present invention also provides a fault prediction device for an energy storage liquid cooling system. Figure 6 The schematic diagram of the structure corresponding to this embodiment is shown. Figure 6The device includes: a data acquisition module 100, which is used to obtain operation monitoring parameters corresponding to a preset energy storage liquid cooling system; the operation monitoring parameters include monitoring parameters corresponding to preset key equipment of the energy storage liquid cooling system; a data processing module 200, which is used to divide the operation monitoring parameters based on the local noise level and time scale change of the operation monitoring parameters to obtain segmented data corresponding to multiple time periods; a calculation module 300, which is used to determine the denoising strength corresponding to the segmented data according to the noise intensity corresponding to the segmented data; a denoising module 400, which is used to denoise the segmented data based on the denoising strength to determine the parameters to be measured corresponding to each time scale of the operation monitoring parameters; an execution module 500, which is used to input the parameters to be measured into a pre-trained neural network model, classify and predict the parameters to be measured through the neural network model, and determine the fault category prediction results corresponding to the parameters to be measured; an output module 600, which is used to determine the fault state corresponding to the energy storage liquid cooling system according to the fault category prediction results.

[0149] An energy storage liquid cooling system fault prediction device provided in an embodiment of the present invention has the same technical features as the above-mentioned method embodiment, and can therefore solve the same technical problems and achieve the same technical effects.

[0150] Furthermore, the above-mentioned execution module 500 is also used to capture the features of the parameters to be measured through the dynamic fusion layer of the time series features of the neural network model, using multiple convolution kernels of different scales, and determine the local target features corresponding to the parameters to be measured; through the adaptive regression unit of the neural network model, using a preset expert selection matrix to adaptively select the activation path of the parameters to be measured, and determine the high-dimensional feature representation corresponding to the parameters to be measured; based on the preset step-size convolution operation, the time scale features corresponding to the local target features and the high-dimensional feature representation are displayed; based on the contribution of the parameters to be measured to the dynamic fusion layer of the time series features or the adaptive regression unit, the time scale features corresponding to the local target features and the time scale features corresponding to the high-dimensional feature representation are adaptively fused to determine the target fusion features corresponding to the parameters to be measured; the fault category probability prediction is performed on the target fusion features to determine the fault category prediction result corresponding to the parameters to be measured.

[0151] The above-mentioned execution module 500 is also used to perform nonlinear mapping on the characteristic parameters of the parameters to be measured at each moment through a preset multi-layer perceptron to determine the query vector corresponding to the parameters to be measured at each moment; determine the convolution kernel attention weight corresponding to the characteristic parameters of the parameters to be measured at each moment according to the attention mechanism of the query vector and the key vector of the preset dynamic convolution kernel; capture the temporal characteristics of the parameters to be measured according to the convolution kernel weight and the time attention weight corresponding to the characteristic parameters, and determine the local target features corresponding to the parameters to be measured; wherein the time attention weight is based on the extraction and determination of the local context information of the parameters to be measured using a time sliding window.

[0152] The above-mentioned execution module 500 is also used to use a preset expert selection matrix to map the parameters to be measured into an expert probability distribution; based on the expert probability distribution, determine the expert units corresponding to the characteristic parameters of the parameters to be measured at each moment; based on the regression parameters corresponding to the expert units, perform nonlinear transformation on the parameters to be measured to determine the high-dimensional feature representation corresponding to the parameters to be measured.

[0153] The above-mentioned data processing module 200 is also used to: determine the data noise difference in the current time period based on the data noise level corresponding to each time point in the preset time period of the operation monitoring parameter; determine the time segmentation point corresponding to the operation monitoring parameter based on the data noise difference; based on the time segmentation point and the preset time segmentation length penalty item, divide the data of the operation monitoring parameter to obtain segmented data corresponding to multiple time scales.

[0154] The above-mentioned execution module 500 is also used to obtain a pre-constructed training sample set; input the training sample set into a preset neural network model in batches, and calculate the classification error, regression error and total loss corresponding to the expert diversity constraint item of the training sample set through a preset hybrid loss function; update the parameters of the neural network model according to the total loss; and, during the parameter update process of the neural network model, based on a preset time-sensitive discarding strategy, randomly discard the feature elements of the training sample set with probability; the probability of random discarding is calculated based on the feature amplitude and diversity corresponding to the feature elements; until the neural network model meets the preset training requirements, select the current optimal weight parameters and generate the target neural network model.

[0155] The above-mentioned execution module 500 is also used to: determine the weighted regression loss weight based on the number of samples corresponding to different sample categories of the preset training sample set; perform parameter space differentiation constraints on the parameter vectors of the preset multiple expert units according to the preset parameter vector constraint algorithm, and determine the expert diversity constraint loss corresponding to the training sample set; wherein the expert unit is used to learn the preset target fault mode characteristics in the training sample set based on the corresponding parameter space; calculate the hybrid loss function corresponding to the training sample set based on the expert diversity constraint loss, the weighted regression loss weight and the preset classification loss.

[0156] The above-mentioned execution module 500 is also used to: obtain the model parameters corresponding to the neural network model in each iteration, and constrain the parameter update amplitude of the neural network model based on the parameter state corresponding to the model parameters; extract the total loss corresponding to the pre-stored memory samples from the pre-set memory library, and update the parameters of the neural network model according to the total loss corresponding to the memory samples; wherein the memory samples are samples in the training sample set.

[0157] The execution module 500 is further configured to: determine new samples corresponding to the training sample set through online learning; measure the similarity between the new samples and the training sample set to determine difference samples between the new samples and the training sample set; and update the memory base based on the difference samples.

[0158] An embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned Figures 1 to 2 The embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to execute the above Figures 1 to 2 The embodiment of the present invention also provides a structural diagram of an electronic device, such as Figure 7 FIG. 1 is a schematic diagram of the structure of the electronic device, wherein the electronic device includes a processor 71 and a memory 70, the memory 70 stores computer executable instructions that can be executed by the processor 71, and the processor 71 executes the computer executable instructions to implement the above Figures 1 to 2 Either of the methods shown. Figure 7 In the illustrated embodiment, the electronic device further includes a bus 72 and a communication interface 73 , wherein the processor 71 , the communication interface 73 and the memory 70 are connected via the bus 72 .

[0159] Among them, the memory 70 may include high-speed random access memory (RAM), and may also include non-volatile memory (non-volatile memory), such as at least one disk storage. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 73 (which can be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 72 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc., or an AMBA (Advanced Microcontroller Bus Architecture, on-chip bus standard) bus, wherein AMBA defines three types of buses, including APB (Advanced Peripheral Bus) bus, AHB (Advanced High-performance Bus) bus and AXI (Advanced eXtensible Interface) bus. The bus 72 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 7 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0160] The processor 71 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by hardware integrated logic circuits in the processor 71 or by software instructions. The above processor 71 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly implemented and executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor 71 reads the information in the memory and combines its hardware to complete the above Figures 1 to 2 Any of the methods shown.

[0161] A computer program product for a method and apparatus for predicting a fault in an energy storage liquid cooling system, provided in an embodiment of the present invention, includes a computer-readable storage medium storing program code. The program code includes instructions that can be used to execute the methods described in the aforementioned method embodiments. For specific implementations, please refer to the method embodiments and will not be described in detail here. Those skilled in the art will clearly understand that, for ease and brevity of description, the specific operating processes of the system described above can refer to the corresponding processes in the aforementioned method embodiments and will not be described in detail here. Furthermore, in the description of the embodiments of the present invention, unless otherwise specified or limited, the terms "installed," "connected," and "connected" should be interpreted broadly. For example, they can refer to fixed, removable, or integral connections; mechanical or electrical connections; direct connections, indirect connections through an intermediary, or internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances. If the functions described are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0162] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0163] Finally, it should be noted that the above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for predicting failure of an energy storage liquid cooling system, characterized in that: The method comprises: Obtaining operation monitoring parameters corresponding to a preset energy storage liquid cooling system; the operation monitoring parameters include monitoring parameters corresponding to preset key equipment of the energy storage liquid cooling system; Based on the local noise level and time scale change of the operation monitoring parameter, the operation monitoring parameter is divided into data to obtain segmented data corresponding to multiple time periods; Determining a denoising strength corresponding to the segmented data according to a noise intensity corresponding to the segmented data; Performing denoising on the segmented data based on the denoising strength, and determining the operation monitoring parameter based on the measured parameter corresponding to each time scale; Inputting the parameters to be measured into a pre-trained neural network model, classifying and predicting the parameters to be measured by the neural network model, and determining the fault category prediction result corresponding to the parameters to be measured; Determining a fault state corresponding to the energy storage liquid cooling system according to the fault category prediction result; The step of dividing the operation monitoring parameters into data based on the local noise level and time scale variation of the operation monitoring parameters to obtain segmented data corresponding to multiple time scales includes: Determining the data noise difference in the current time period based on the data noise level corresponding to each time point of the preset time period of the operation monitoring parameter; Determining time segment points corresponding to the operation monitoring parameters based on the data noise difference; Based on the time segmentation points and the preset time segmentation length penalty item, the operation monitoring parameters are divided into data to obtain segmented data corresponding to multiple time scales.

2. The method according to claim 1, characterized in that The step of classifying and predicting the parameters to be measured by the neural network model to determine the fault category prediction results corresponding to the parameters to be measured includes: The time series feature dynamic fusion layer of the neural network model is used to capture the features of the parameters to be measured using multiple convolution kernels of different scales to determine the local target features corresponding to the parameters to be measured; By using the adaptive regression unit of the neural network model, a preset expert selection matrix is ​​used to adaptively select an activation path for the parameter to be measured, and a high-dimensional feature representation corresponding to the parameter to be measured is determined; Based on a preset step-length convolution operation, the local target features and the high-dimensional feature representations are displayed to represent the time scale features corresponding to each other; Based on the contribution of the parameter to be measured to the dynamic fusion layer of the temporal features or the adaptive regression unit, adaptively fuse the time scale features corresponding to the local target features and the time scale features corresponding to the high-dimensional feature representation to determine the target fusion feature corresponding to the parameter to be measured; Perform fault category probability prediction on the target fusion feature to determine the fault category prediction result corresponding to the parameter to be measured.

3. The method according to claim 2, characterized in that The steps of capturing the features of the parameters to be measured by using a plurality of convolution kernels of different scales through the dynamic fusion layer of the time series features of the neural network model and determining the local target features corresponding to the parameters to be measured include: Performing nonlinear mapping on the characteristic parameters of the parameter to be measured at each moment through a preset multi-layer perceptron to determine the query vector corresponding to the parameter to be measured at each moment; Determine the convolution kernel attention weight corresponding to the feature parameter of the parameter to be measured at each moment according to the attention mechanism of the query vector and the key vector of the preset dynamic convolution kernel; According to the convolution kernel attention weight and the time attention weight corresponding to the feature parameter, the temporal characteristics of the parameter to be measured are captured and the local target features corresponding to the parameter to be measured are determined; wherein, the time attention weight is determined based on the extraction of the local context information of the parameter to be measured using a time sliding window.

4. The method according to claim 2, characterized in that The steps of adaptively selecting activation paths for the parameters to be measured using a preset expert selection matrix by the adaptive regression unit of the neural network model and determining high-dimensional feature representations corresponding to the parameters to be measured include: Mapping the parameters to be measured into expert probability distribution using a preset expert selection matrix; Determine, based on the expert probability distribution, the expert units corresponding to the characteristic parameters of the parameter to be measured at each moment; Based on the regression parameters corresponding to the expert units, a nonlinear transformation is performed on the parameters to be measured to determine the high-dimensional feature representation corresponding to the parameters to be measured.

5. The method according to claim 1, wherein The method further comprises: Obtain a pre-built training sample set; Inputting the training sample set into a preset neural network model in batches, and calculating the classification error, regression error and total loss corresponding to the expert diversity constraint term of the training sample set using a preset hybrid loss function; updating the parameters of the neural network model according to the total loss; and, during the parameter updating process of the neural network model, randomly discarding the characteristic elements of the training sample set based on a preset time-sensitive discarding strategy; the probability of random discarding is calculated based on the characteristic amplitude and diversity corresponding to the characteristic elements; Until the neural network model meets the preset training requirements, the current optimal weight parameters are selected to generate the target neural network model.

6. The method according to claim 5, characterized in that The calculation method of the loss function includes: Determine the weighted regression loss weight according to the number of samples corresponding to different sample categories in the preset training sample set; According to a preset parameter vector constraint algorithm, parameter space differentiation constraints are respectively applied to the parameter vectors of the preset multiple expert units to determine the expert diversity constraint loss corresponding to the training sample set; wherein the expert units are used to learn the preset target fault mode characteristics in the training sample set based on the corresponding parameter space; A hybrid loss function corresponding to the training sample set is calculated according to the expert diversity constraint loss, the weighted regression loss weight and the preset classification loss.

7. The method according to claim 5, characterized in that The step of updating the parameters of the neural network model according to the total loss includes: Obtaining model parameters corresponding to each iteration of the neural network model, and constraining the parameter update amplitude of the neural network model based on parameter states corresponding to the model parameters; The total loss corresponding to the pre-stored memory samples is extracted from a pre-set memory library, and the parameters of the neural network model are updated according to the total loss corresponding to the memory samples; wherein the memory samples are samples in the training sample set.

8. The method according to claim 7, characterized in that The method further comprises: Determining new samples corresponding to the training sample set through online learning; Measuring the similarity between the new sample and the training sample set, and determining the difference samples between the new sample and the training sample set; Based on the difference samples, data in the memory bank is updated.

9. A fault prediction device for an energy storage liquid cooling system, characterized in that: The device comprises: A data acquisition module, configured to acquire operation monitoring parameters corresponding to a preset energy storage liquid cooling system; the operation monitoring parameters include monitoring parameters corresponding to preset key equipment of the energy storage liquid cooling system; a data processing module, configured to divide the operation monitoring parameters into data segments based on the local noise level and time scale variation of the operation monitoring parameters, and obtain segmented data corresponding to a plurality of time periods; a calculation module, configured to determine a denoising strength corresponding to the segmented data according to a noise intensity corresponding to the segmented data; a denoising module, configured to perform denoising on the segmented data based on the denoising strength, and determine the operation monitoring parameter based on the measured parameter corresponding to each time scale; An execution module, configured to input the parameters to be measured into a pre-trained neural network model, perform classification prediction on the parameters to be measured through the neural network model, and determine a fault category prediction result corresponding to the parameters to be measured; An output module, configured to determine a fault state corresponding to the energy storage liquid cooling system according to the fault category prediction result; The data processing module is also used to: determine the data noise difference in the current time period based on the data noise level corresponding to each time point of the preset time period of the operation monitoring parameter; determine the time segmentation point corresponding to the operation monitoring parameter based on the data noise difference; and divide the data of the operation monitoring parameter based on the time segmentation point and the preset time segment length penalty item to obtain segmented data corresponding to multiple time scales.

Images