A method and terminal for abnormal early warning of DC / DC devices based on the albert_tiny model

Abstract

This invention discloses a method and terminal for abnormal warning of DC / DC devices. The method involves acquiring the input power, power efficiency, and capacitor temperature of a DC / DC device at a certain moment as a data acquisition unit; integrating the data acquisition units (including a first preset number) to generate a feature sequence as a sample; constructing a running sequence dataset using a second preset number of samples; pre-training a deep learning model using the running sequence data; fine-tuning the deep learning model using a running sequence dataset labeled with abnormal data to generate an abnormal warning judgment model; acquiring the sample consisting of the current DC / DC device's input power, power efficiency, and capacitor temperature as the input to the abnormal warning judgment model; and determining whether an alarm needs to be issued based on the output of the abnormal warning judgment model.

Description

[0001] This case is a divisional application based on the invention patent filed on May 29, 2024, with application number 202410675781.8 and titled "A Method and Terminal for Early Warning of Abnormalities in DC / DC Devices". Technical Field

[0002] This invention relates to the field of energy storage, and in particular to a method and terminal for early warning of abnormalities in DC / DC equipment. Background Technology

[0003] With the increasing market share of new energy vehicles and the widespread adoption of supercharging and fast charging, the application of photovoltaic-storage charging and testing stations is gradually becoming more common. Among these stations, DC / DC converters (direct current converter components) are key equipment that requires frequent maintenance and replacement; sudden damage can severely disrupt the normal charging order of the stations.

[0004] Traditional DC / DC fault warning systems primarily monitor the DC / DC voltage and current, setting thresholds. During daily operation, alarms are triggered when these thresholds are exceeded. This approach has a high false alarm rate, easily causing unnecessary downtime and increasing operating costs. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and terminal for early warning of abnormalities in DC / DC equipment, so as to accurately detect DC / DC equipment faults in advance.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A method for early warning of abnormalities in DC / DC devices includes the following steps:

[0008] S1. Acquire the input power, power efficiency, and capacitor temperature of the DC / DC device at a certain moment as a collection unit, integrate the collection units including a first preset number, and generate a feature sequence as a sample; the sample is a three-dimensional feature sequence composed of the input power, power efficiency, and capacitor temperature in the first preset number of collection units.

[0009] S2. Construct a running sequence dataset using the second preset number of samples; train the albert_tiny model using the running sequence dataset and return a deep learning model that can represent sequence relationships and sequence states; the deep learning model uses a fully connected layer to replace the original input embedding layer;

[0010] S3. Construct a dataset of running sequences with anomaly data annotations to fine-tune the deep learning model and generate an anomaly warning and judgment model;

[0011] S4. Obtain the sample consisting of the input power, power efficiency, and capacitor temperature of the current DC / DC device as the input of the anomaly warning determination model, and determine whether an alarm needs to be issued based on the output of the anomaly warning determination model.

[0012] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows:

[0013] A terminal for abnormal early warning of DC / DC equipment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps:

[0014] S1. Acquire the input power, power efficiency, and capacitor temperature of the DC / DC device at a certain moment as a collection unit, integrate the collection units including a first preset number, and generate a feature sequence as a sample; the sample is a three-dimensional feature sequence composed of the input power, power efficiency, and capacitor temperature in the first preset number of collection units.

[0015] S2. Construct a running sequence dataset using the second preset number of samples; train the albert_tiny model using the running sequence dataset and return a deep learning model that can represent sequence relationships and sequence states; the deep learning model uses a fully connected layer to replace the original input embedding layer;

[0016] S3. Construct a dataset of running sequences with anomaly data annotations to fine-tune the deep learning model and generate an anomaly warning and judgment model;

[0017] S4. Obtain the sample consisting of the input power, power efficiency, and capacitor temperature of the current DC / DC device as the input of the anomaly warning determination model, and determine whether an alarm needs to be issued based on the output of the anomaly warning determination model.

[0018] The beneficial effects of this invention are as follows: It provides a method and terminal for early warning of abnormalities in DC / DC devices, which obtains the input power, power efficiency and capacitor temperature of the DC / DC device during operation as the basic acquisition unit for constructing samples, effectively captures the operating status of the device, forms meaningful samples, and uses these samples to pre-train and fine-tune a deep learning model to construct an efficient abnormality early warning judgment model, thereby discovering potential equipment failures in advance and avoiding greater economic losses and safety accidents. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for early warning of abnormalities in DC / DC devices according to an embodiment of the present invention;

[0020] Figure 2 This is a flowchart illustrating a method for early warning of abnormalities in DC / DC devices according to an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the architecture of the anomaly warning determination model in a method for anomaly warning of DC / DC equipment according to an embodiment of the present invention;

[0022] Figure 4 This is a schematic diagram of a DC / DC device anomaly warning terminal according to an embodiment of the present invention;

[0023] Figure 5 This is a topology diagram of the ALBERT_Tiny model in this embodiment of the invention;

[0024] Label Explanation:

[0025] 1. A terminal for abnormal early warning of DC / DC equipment; 2. Memory; 3. Processor; Detailed Implementation

[0026] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0027] The relevant terms used in this invention are explained below:

[0028] DC / DC converter: A device that converts the voltage of an input DC power supply into different output DC voltages.

[0029] Pre-trained models: Pre-trained models are neural network models that have been pre-trained on large-scale data using unsupervised or semi-supervised learning methods. These models are typically trained using powerful computing resources and large amounts of text, image, or audio data to learn the statistical structure and feature representations of the data. The purpose of pre-trained models is to extract universal feature representations by learning patterns and information from large-scale data, so that they can be fine-tuned or used for transfer learning in subsequent tasks, thereby improving the model's performance on specific tasks.

[0030] ALBERT_Tiny: A smaller version of the ALBERT (ALite BERT) model. ALBERT is an improved pre-trained model based on BERT (Bidirectional Encoder Representations from Transformers), developed by the Google Brain team.

[0031] Embedding: Embedding is an important concept in computer science and machine learning. It generally refers to the process of mapping high-dimensional data to a low-dimensional space, where each data point (e.g., a word, image, user, etc.) is represented as a vector, which typically contains fewer dimensions. It maps discrete symbolic representations (such as words, characters, categories, etc.) to a continuous space of real-valued vectors. Embedding is often used to transform high-dimensional, sparse, discrete features into low-dimensional, dense, continuous features, thereby better expressing the semantic relationships between features.

[0032] Fully Connected Layer: Also known as a densely connected layer or simply a fully connected layer, a fully connected layer is a common layer type in deep learning neural networks. In a fully connected layer, each node (neuron) is connected to all nodes in the previous layer. Therefore, each input feature is multiplied by the weights of each neuron and then passed through an activation function to obtain the output.

[0033] Layer normalization is a normalization technique used in neural networks to address problems such as vanishing and exploding gradients during deep neural network training. In layer normalization, the features of each sample are normalized so that the mean of each feature in the sample dimension is 0 and the standard deviation is 1.

[0034] Dropout: Dropout is a regularization technique used in deep neural networks to prevent overfitting. During training, Dropout randomly sets some neurons in the network to 0, essentially "dropping" them. During forward and backward propagation, the activation values ​​of these neurons are set to 0. This means that the model is randomly pruned in each training iteration, reducing model complexity to some extent.

[0035] Transformer: Transformer is a deep learning model used to process sequential data. The Transformer model consists of an encoder and a decoder, and is often used to handle sequence-to-sequence (Seq2Seq) tasks.

[0036] Softmax: Softmax is a commonly used activation function, mainly used for the output of the last layer in multi-class classification problems.

[0037] Please refer to Figure 1 A method for early warning of abnormalities in DC / DC devices, comprising the following steps:

[0038] S1. Acquire the input power, power efficiency, and capacitor temperature of the DC / DC device at a certain moment as a collection unit, integrate the collection units including a first preset number, and generate a feature sequence as a sample; the sample is a three-dimensional feature sequence composed of the input power, power efficiency, and capacitor temperature in the first preset number of collection units.

[0039] S2. Construct a running sequence dataset using the second preset number of samples; train the albert_tiny model using the running sequence dataset and return a deep learning model that can represent sequence relationships and sequence states; the deep learning model uses a fully connected layer to replace the original input embedding layer;

[0040] S3. Construct a dataset of running sequences with anomaly data annotations to fine-tune the deep learning model and generate an anomaly warning and judgment model;

[0041] S4. Obtain the sample consisting of the input power, power efficiency, and capacitor temperature of the current DC / DC device as the input of the anomaly warning determination model, and determine whether an alarm needs to be issued based on the output of the anomaly warning determination model.

[0042] It is understandable that the running data series dataset with abnormal data annotations is a dataset formed when abnormal conditions occur in the application context of DC / DC equipment, corresponding to the input power, power efficiency, and capacitor temperature.

[0043] As described above, the beneficial effects of this invention are as follows: It acquires the input power, power efficiency, and capacitor temperature of the DC / DC device during operation as the basic data collection units for constructing samples, effectively capturing the device's operating status and forming meaningful samples. These samples are then used for pre-training and fine-tuning of deep learning models, constructing an efficient anomaly early warning and judgment model, thereby detecting potential equipment failures in advance and avoiding greater economic losses and safety accidents.

[0044] Please refer to Figures 1 to 3 In an embodiment of the present invention, the deep learning model is the ALBERT_Tiny model, and a fully connected layer is used to replace the original input embedding layer;

[0045] Step S2 specifically involves:

[0046] A running sequence dataset is constructed using the second preset number of samples;

[0047] The ALBERT_Tiny model is trained using the aforementioned running sequence dataset, and a deep learning model capable of representing sequence relationships and sequence states is returned.

[0048] As described above, replacing the original embedding layer with a fully connected layer using the ALBERT_Tiny model reduces model complexity and improves computational efficiency, while maintaining the model's high sensitivity to sequence relationships and states. This improvement helps enhance the model's performance when processing time series data, especially in identifying abnormal device behavior.

[0049] In an embodiment of the present invention, step S3 specifically comprises:

[0050] Freeze some parameters of the deep learning model;

[0051] Construct a dataset of running sequences labeled with anomalous data;

[0052] The deep learning model is fine-tuned using the runtime sequence dataset with anomaly data annotations to generate an anomaly warning and judgment model.

[0053] As described above, by freezing some parameters of a deep learning model and fine-tuning it using data labeled with anomalies, the model can be made to focus more on learning features that distinguish between normal and abnormal states. This method improves the model's generalization ability and accuracy in practical applications, ensuring the reliability of anomaly warnings.

[0054] In an embodiment of the present invention, the sample is a three-dimensional feature sequence composed of input power, power efficiency, and capacitor temperature from a first preset number of acquisition units.

[0055] As can be seen from the above description, clearly defining the three-dimensional feature sequence of the sample as consisting of input power, power efficiency, and capacitor temperature enables the model to comprehensively evaluate the operating status of the equipment, thereby enhancing the comprehensive judgment capability of the early warning system.

[0056] Specifically, the sample structure is exemplified as follows:

[0057] [[P1,P2,…,P 64 ],[E1,E2,…,E 64 ],[T1,T2,…,T 64 ]]

[0058] Wherein, the input power is denoted as P n Power efficiency is denoted as E n The capacitor temperature is denoted as T. n 64 points were taken in chronological order of sampling time to form a 64*3-dimensional feature sequence.

[0059] In an embodiment of the present invention, the first preset quantity is equal to the second preset quantity and the value range is 50-100.

[0060] As can be seen from the above description, based on the application background of optical storage charging and testing stations, experience shows that when the first preset quantity and the second preset quantity are equal and the value range is 50-100, the model training effect is the best, preferably 60 and 64.

[0061] Please refer to Figure 4 A terminal for abnormal warning of DC / DC equipment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it completes the steps in the above-mentioned method for abnormal warning of DC / DC equipment.

[0062] As described above, this invention provides an execution platform for a method of early warning of anomalies in DC / DC devices. When executing this method, the input power, power efficiency, and capacitor temperature of the DC / DC device during operation are acquired as the basic data collection units for constructing samples. This effectively captures the operating status of the device, forming meaningful samples. These samples are then used for pre-training and fine-tuning of a deep learning model to construct an efficient anomaly warning and judgment model, thereby detecting potential device failures in advance and avoiding greater economic losses and safety accidents.

[0063] This invention provides a method and terminal for early warning of DC / DC equipment anomalies, mainly used for accurate early warning of anomalies in DC / DC equipment in optical storage charging and testing stations. The following is a detailed description with reference to embodiments:

[0064] Please refer to Figure 1 Embodiment 1 of the present invention is as follows:

[0065] A method for early warning of abnormalities in DC / DC devices includes the following steps:

[0066] S1. Acquire the input power, power efficiency, and capacitor temperature of the DC / DC device at a certain moment as a collection unit, integrate the collection units containing a first preset number, and generate a feature sequence as a sample;

[0067] S2. Construct a running sequence dataset using a second preset number of samples and pre-train the deep learning model using the running sequence data solution;

[0068] S3. Construct a dataset of running sequences with anomaly data annotations to fine-tune the deep learning model and generate an anomaly warning and judgment model;

[0069] S4. Obtain a sample consisting of the input power, power efficiency, and capacitor temperature of the current DC / DC device as input to the anomaly warning judgment model, and determine whether an alarm needs to be issued based on the output of the anomaly warning judgment model.

[0070] In this embodiment, the input power, power efficiency, and capacitor temperature of the DC / DC device during operation are used as the basic data acquisition units for constructing samples, effectively capturing the device's operating status and forming meaningful samples. These samples are then used for pre-training and fine-tuning of a deep learning model to build an efficient anomaly early warning and judgment model, thereby detecting potential equipment failures in advance and avoiding greater economic losses and safety accidents.

[0071] Specifically, step S0 is included before step S1:

[0072] Obtain the input power and output power of the DC / DC device at a certain moment, and calculate the power efficiency.

[0073] Specifically, the fine-tuning principle and process in step S3 are as follows:

[0074] A pre-trained-fine-tuned model is used, where:

[0075] Pre-training: In this stage, a model is trained using a large-scale unlabeled dataset (usually text data). This model is typically a deep neural network, such as the Transformer model. During training, the model captures abstract representations of language by learning language patterns, grammatical structures, and semantic information from the text data. The purpose of pre-training is to provide the model with a general understanding of the linguistic world.

[0076] Fine-tuning: After pre-training, the pre-trained model is used for a specific task (such as text classification, named entity recognition, sentiment analysis, etc.). In the fine-tuning phase, the model is trained on a small amount of labeled data to adjust its parameters, making it more suitable for the specific task. The fine-tuning process typically involves fixing some or all of the pre-trained model's parameters and then updating the remaining parameters on new labeled data.

[0077] The advantage of the pre-trained-fine-tuned model training method lies in its ability to utilize large-scale unlabeled data for pre-training, thereby giving the model better generalization ability. Furthermore, it requires only a relatively small amount of labeled data for fine-tuning on specific tasks, saving on data labeling costs. In addition, since the pre-trained model has already learned the general rules of language, it often achieves better performance on specific tasks.

[0078] Please refer to Figures 1 to 3 Embodiment two of the present invention is as follows:

[0079] Based on Example 1, the deep learning model is the ALBERT_Tiny model, and a fully connected layer is used to replace the original input embedding layer. Using the ALBERT_Tiny model and replacing the original embedding layer with a fully connected layer can reduce the complexity of the model and improve computational efficiency, while maintaining the model's high sensitivity to sequence relationships and states.

[0080] Specifically, please refer to Figure 5 The ALBERT_Tiny model includes Timedistributed(dense), which means that a dense operation is performed on each of the 64 time steps in the sequence. Here, Dense, or Dense Layer, also known as a Fully Connected Layer or Affine Layer, is specifically expressed as:

[0081] Y = f(W*X + b);

[0082] Where X represents the input data, with a shape of (n, m), where n is the number of samples and m is the dimension of the input features;

[0083] W represents the connection weight matrix, with shape (m, k). ij This represents the connection weight between the i-th input feature and the j-th neuron;

[0084] b is the bias vector with shape (1, k);

[0085] f is the activation function applied to XW, such as ReLU, Sigmoid, etc.

[0086] Y represents the output;

[0087] Specifically, the ALBERT_Tiny model operates as follows:

[0088] (1) Input data:

[0089] Input power sequence (1*64)

[0090] Power efficiency sequence (1*64)

[0091] Capacitor temperature sequence (1*64)

[0092] These input data represent the input power, power efficiency, and capacitor temperature of the DC-DC device, respectively, and are input into the model in the form of time series, with each series containing 64 time steps of values.

[0093] (2) Merging sequences:

[0094] The three input sequences are merged into a single sequence (3*64), which means that the data from each time step are spliced ​​together to form a time series containing three features.

[0095] (3) Position Embedding:

[0096] Add position embedding information (128*64) to each time step of the merged sequence. Through position embedding, the model can identify the position of each data point in the time series, thereby better understanding the sequential information of the time series.

[0097] (4) TimeDistributed(Dense)Layer:

[0098] A full connection (128*64) is performed on each time step of the merged sequence, that is, a full connection operation is performed on the data of each time step to extract features and increase the feature dimension.

[0099] (5) Plus (adding embedded elements):

[0100] The time series output by the fully connected layer is added to the location embedding vector to further fuse the location information of the time step.

[0101] (6) LayerNorm and Dropout:

[0102] The data is standardized by layer normalization, followed by Dropout operation to randomly discard some neurons, in order to prevent overfitting and improve the model's generalization ability.

[0103] (7) Transformer (transformer module):

[0104] A 12-layer transformer module is used, with a hidden size of 768 for each layer and 12 `num_hidden_layers`. The transformer module performs feature extraction and sequence modeling on time series data through a multi-head attention mechanism and a feedforward neural network.

[0105] Step S2 is as follows:

[0106] Construct a running sequence dataset using a second preset number of samples;

[0107] The ALBERT_Tiny model is trained using a running sequence dataset, and a deep learning model capable of representing sequence relationships and states is returned. These steps help improve the model's performance when processing time series data, especially in identifying abnormal device behavior.

[0108] Specifically, refer to Figure 3 Step S3 is as follows:

[0109] Freezing some parameters of a deep learning model; by freezing some parameters of a deep learning model and fine-tuning it using data with anomaly labels, the model can be made to focus more on learning features that distinguish between normal and abnormal states.

[0110] Construct a dataset of running sequences labeled with anomalous data;

[0111] The deep learning model was fine-tuned using a dataset of runtime sequences labeled with anomalies, generating an anomaly warning and judgment model. These steps improved the model's generalization ability and accuracy in practical applications, ensuring the reliability of anomaly warnings.

[0112] Specifically, in this example, the early warning system for DC / DC devices is implemented in two parts: online and offline. The online part uses a trained anomaly warning judgment model to provide real-time warnings for DC / DC devices. The offline part trains an anomaly warning judgment model based on historical data using the above method. The model is continuously updated by replacing samples with daily data from DC / DC devices to improve the accuracy of the model's judgment.

[0113] Embodiment 3 of the present invention is as follows:

[0114] Based on Example 2, the sample structure is as follows:

[0115] [[P1,P2,…,P 64 ],[E1,E2,…,E 64 ],[T1,T2,…,T 64 ]]

[0116] Wherein, the input power is denoted as P n Power efficiency is denoted as E n The capacitor temperature is denoted as T. n 64 points were taken in chronological order of sampling time to form a 64*3-dimensional feature sequence.

[0117] Please refer to Figure 4 Embodiment four of the present invention is as follows:

[0118] A terminal 1 for abnormal warning of DC / DC equipment includes a memory 2, a processor 3, and a computer program stored in the memory 2 and executable on the processor 3. When the processor 3 executes the computer program, it completes the steps in any of the DC / DC equipment abnormal warning methods described in embodiments one to three above.

[0119] In summary, the present invention provides a method and terminal for abnormal early warning of DC / DC equipment. It acquires the input power, power efficiency and capacitor temperature of the DC / DC equipment during operation as the basic acquisition unit for constructing samples, effectively captures the operating status of the equipment, forms meaningful samples, and uses these samples to pre-train and fine-tune a deep learning model to construct an efficient abnormal early warning judgment model, thereby discovering potential equipment failures in advance and avoiding greater economic losses and safety accidents.

[0120] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for early warning of anomalies in DC / DC devices based on the Albert_tiny model, characterized in that: Including the following steps: S1. Acquire the input power, power efficiency, and capacitor temperature of the DC / DC device at a certain moment as a collection unit, integrate the collection units including a first preset number, and generate a feature sequence as a sample; the sample is a three-dimensional feature sequence composed of the input power, power efficiency, and capacitor temperature in the first preset number of collection units. S2. Construct a running sequence dataset using the second preset number of samples; train the albert_tiny model using the running sequence dataset and return a deep learning model that can represent sequence relationships and sequence states; the deep learning model uses a fully connected layer to replace the original input embedding layer; S3. Construct a dataset of running sequences with anomaly data annotations to fine-tune the deep learning model and generate an anomaly warning and judgment model; S4. Obtain the sample consisting of the input power, power efficiency, and capacitor temperature of the current DC / DC device as the input of the anomaly warning determination model, and determine whether an alarm needs to be issued based on the output of the anomaly warning determination model. The operation of the albert_tiny model includes the following steps: Input power sequence, power efficiency sequence, and capacitor temperature sequence of the input DC / DC device; Merge the three input sequences into a single sequence; Add location embedding information to each time series in the merged sequence; Perform a full connection on each time series of the merged sequence; Add the time series output by the fully connected layer to the location embedding vector; The data is standardized by layer normalization, followed by dropout. Complex feature extraction is performed using Transformer units.

2. The method for early warning of DC / DC device anomalies based on the albert_tiny model according to claim 1, characterized in that: The albert_tiny model specifically includes: Input unit: used to receive power sequence, power efficiency sequence and capacitor temperature sequence, each sequence having a dimension of 1×n; where n is a positive integer; The data merging unit is used to merge the power sequence, power efficiency sequence, and capacitor temperature sequence into a 3×n matrix; where n is a positive integer. Location embedding unit, used to add location label to each data point in the merged data sequence; The temporal distribution unit includes a fully connected layer for performing feature transformation on each time point of the merged sequence, mapping the data at each time point to a 128-dimensional space. The addition unit is used to add the feature-transformed data with the location embedding information to fuse the location information; The normalization and dropout units are used to normalize the data and randomly remove some data to prevent overfitting. The Transformer unit consists of 12 layers, each containing 768 hidden units, and is used to capture the relationship between distant data points in a sequence through a self-attention mechanism to extract complex features. Step S3 specifically involves: Freeze some parameters of the deep learning model; Construct a dataset of running sequences labeled with anomalous data; The deep learning model is fine-tuned using the runtime sequence dataset with anomaly data annotations to generate an anomaly warning and judgment model.

3. The method for early warning of DC / DC device anomalies based on the albert_tiny model according to claim 1, characterized in that: The structure of the sample is as follows: ； Wherein, the input power is denoted as P n Power efficiency is denoted as E n The capacitor temperature is denoted as T. n The n is a positive integer.

4. The method for early warning of DC / DC device anomalies based on the albert_tiny model according to claim 2, characterized in that: The temporal distribution unit includes a fully connected layer with a temporal distribution, used to perform independent feature transformation on the data at each time point of the merged sequence; The Transformer unit extracts features through a self-attention mechanism to capture the relationships between distant data points in the sequence; The normalization and dropout units are used to standardize the data after feature transformation and position fusion, and randomly remove a portion of the data.

5. A terminal for abnormal early warning of DC / DC devices based on the Albert_tiny model, characterized in that: Includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the following steps: S1. Acquire the input power, power efficiency, and capacitor temperature of the DC / DC device at a certain moment as a collection unit, integrate the collection units including a first preset number, and generate a feature sequence as a sample; the sample is a three-dimensional feature sequence composed of the input power, power efficiency, and capacitor temperature in the first preset number of collection units. S2. Construct a running sequence dataset using the second preset number of samples; train the albert_tiny model using the running sequence dataset and return a deep learning model that can represent sequence relationships and sequence states; the deep learning model uses a fully connected layer to replace the original input embedding layer; S3. Construct a dataset of running sequences with anomaly data annotations to fine-tune the deep learning model and generate an anomaly warning and judgment model; S4. Obtain the sample consisting of the input power, power efficiency, and capacitor temperature of the current DC / DC device as the input of the anomaly warning determination model, and determine whether an alarm needs to be issued based on the output of the anomaly warning determination model. The operation of the albert_tiny model includes the following steps: Input power sequence, power efficiency sequence, and capacitor temperature sequence of the DC-DC device; Merge the three input sequences into a single sequence; Add location embedding information to each time series in the merged sequence; Perform a full connection on each time series of the merged sequence; Add the time series output by the fully connected layer to the location embedding vector; The data is standardized by layer normalization, followed by dropout. Complex feature extraction is performed using Transformer units.

6. A terminal for abnormal early warning of DC / DC devices based on the albert_tiny model according to claim 5, characterized in that: The albert_tiny model specifically includes: Input unit: Used to receive power sequence, power efficiency sequence, and capacitor temperature sequence, each sequence having a dimension of 1×n, where n is a positive integer. The data merging unit is used to merge the power sequence, power efficiency sequence, and capacitor temperature sequence into a 3×n matrix, where n is a positive integer. Location embedding unit, used to add location label to each data point in the merged data sequence; The temporal distribution unit includes a fully connected layer for performing feature transformation on each time point of the merged sequence, mapping the data at each time point to a 128-dimensional space. The addition unit is used to add the feature-transformed data with the location embedding information to fuse the location information; The normalization and dropout units are used to normalize the data and randomly remove some data to prevent overfitting. The Transformer unit consists of 12 layers, each containing 768 hidden units, and is used to capture the relationship between distant data points in a sequence through a self-attention mechanism to extract complex features. Step S3 specifically involves: Freeze some parameters of the deep learning model; Construct a dataset of running sequences labeled with anomalous data; The deep learning model is fine-tuned using the runtime sequence dataset with anomaly data annotations to generate an anomaly warning and judgment model.

7. A terminal for abnormal early warning of DC / DC devices based on the albert_tiny model according to claim 5, characterized in that: The structure of the sample is as follows: ； Wherein, the input power is denoted as P n Power efficiency is denoted as E n The capacitor temperature is denoted as T. n The n is a positive integer.

8. A terminal for abnormal early warning of DC / DC devices based on the albert_tiny model according to claim 6, characterized in that: The temporal distribution unit includes a fully connected layer with a temporal distribution, used to perform independent feature transformation on the data at each time point of the merged sequence; The Transformer unit extracts features through a self-attention mechanism to capture the relationships between distant data points in the sequence; The normalization and dropout units are used to standardize the data after feature transformation and position fusion, and randomly remove a portion of the data.

Images