Electric power work order visual management early warning detection system and method

Abstract

The invention discloses an electric power work order visual management early warning detection system and method, and the system comprises a data preprocessing module which is used for preprocessing historical work order data; the first model training module is used for training a first model based on the preprocessed historical work order data to extract features; the feature fusion module is used for fusing features extracted by the first model to construct a prediction model; the real-time data processing module is used for inputting the real-time data into the prediction model, extracting the most probable early warning label and calculating a probability difference; the result output and model optimization module is used for outputting a result and iteratively optimizing a prediction model according to real-time data; wherein the prediction model adopts a uniflow model to fuse text, audio and structured features, and optimization is carried out through a Transform layer and cross entropy loss. According to the system, the data quality is improved, text semantics are accurately captured, time sequence and local features are mined, real-time early warning is achieved, a model is dynamically optimized, and efficient operation and maintenance of power enterprises are assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of work order management and early warning, and more specifically, to a power work order visualization management and early warning detection system and method. Background Art

[0002] In the power industry, the work order management system is a key tool to ensure the stable operation and efficient maintenance of power facilities. Power work orders cover various types of information such as equipment failures, maintenance tasks, and customer services. The data volume is huge and complex, including text descriptions, audio records, and structured data in various modalities. With the continuous expansion and intelligent development of the power system, higher requirements are put forward for the processing and analysis of work order data. Especially in the aspect of early warning detection, it is necessary to be able to extract key information from a large amount of work order data in a timely and accurate manner, predict potential failure risks, so as to take measures in advance to reduce power outage events and maintenance costs.

[0003] Traditional work order processing methods mainly rely on manual review and experience judgment. This method is inefficient and prone to omissions and misjudgments. In recent years, with the development of artificial intelligence and machine learning technologies, data-driven early warning detection methods have gradually emerged. These methods construct prediction models to learn and analyze historical work order data, so as to achieve automatic early warning of new work orders. However, there are some limitations in the existing technologies when dealing with multi-modal work order data. For example, the semantic understanding of text data is not deep enough, the feature extraction of audio data is not accurate enough, and the integration between structured data and unstructured data is not sufficient, resulting in the prediction performance of the model not reaching the ideal state.

[0004] The publication number is CN108306286A, and the name is a method and device for managing active warning work orders in a distribution network, which relates to the technical field of power system regulation and operation. By generating active warning work orders and tracking and monitoring the completion status of active warning work orders, the closed-loop management and control of data abnormal work orders are realized, and the repeated processing of work orders is avoided.

[0005] The publication number is CN117455177A, and the name is a method, device, electronic device, and storage medium for power service early warning. By applying artificial intelligence technology in a power-related system, the acceptance time of work orders is obtained in batches according to preset conditions, and the overdue work order reminder and early warning are automatically calculated according to the current time and the time limit requirements of business expansion and installation.

[0006] However, power work order data has the characteristics of time series and locality. Fault phenomena may gradually appear over a period of time or may break out concentratedly in local equipment or areas. Existing technologies often fail to fully exploit these time series and local characteristics and are unable to accurately capture the development trends and correlation patterns of faults. Moreover, in terms of model optimization, there is a lack of an effective real-time feedback mechanism, and the model parameters cannot be adjusted in a timely manner according to new work order data, resulting in poor adaptability and robustness of the model.

[0007] Therefore, there is an urgent need for a new visual management early warning detection system for power work orders, which can comprehensively consider the characteristics of multi-modal data, deeply explore time series and local characteristics, and achieve dynamic optimization of the model through real-time data feedback, so as to improve the accuracy and timeliness of early warning detection and provide more powerful guarantee for the stable operation of the power system. Therefore, how to provide one is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] In view of this, the present invention provides a visual management early warning detection system and method for power work orders, aiming to solve the technical problems that fault phenomena may gradually appear over a period of time or may break out concentratedly in local equipment or areas. Existing technologies often fail to fully exploit these time series and local characteristics and are unable to accurately capture the development trends and correlation patterns of faults.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] A visual management early warning detection method for power work orders, comprising:

[0011] A data preprocessing module, configured to preprocess historical work order data;

[0012] A first model training module, configured to train a first model based on the preprocessed historical work order data to extract features;

[0013] A feature fusion module, configured to fuse the features extracted by the first model to construct a prediction model;

[0014] A real-time data processing module, configured to input real-time data into the prediction model, extract early warning labels, and calculate the probability difference;

[0015] A result output and model optimization module, configured to output results and iteratively optimize the prediction model according to real-time data;

[0016] Wherein, the prediction model adopts a single-stream model to fuse text, audio, and structured features, and is optimized through a Transformer layer and cross-entropy loss.

[0017] Further, the data preprocessing module includes a data cleaning unit, a word segmentation unit, and a noise reduction unit;

[0018] The data cleaning unit is used to remove error data, duplicate data, and fill in missing data;

[0019] The word segmentation unit is used to select a word segmentation tool and dictionary, and perform word segmentation operations;

[0020] The noise reduction unit is used for text noise reduction, audio noise reduction, and structured data noise reduction.

[0021] Furthermore, the first model training module uses the Word2Vec model to perform word vector training on text data to obtain the feature representation of the text, including a model parameter setting unit, a word vector training unit, and an evaluation and optimization unit;

[0022] The model parameter setting unit is used to determine the word vector dimension, set the window size, and select a training algorithm;

[0023] The word vector training unit is used to initialize the word vector, perform forward propagation calculation, calculate the loss function, perform backpropagation update, and iterative training;

[0024] The evaluation and optimization unit is used to evaluate the quality of the word vector and adjust the model parameters.

[0025] Furthermore, the feature fusion module includes a temporal information learning unit and a local feature extraction unit;

[0026] The temporal information learning unit is used to perform temporal information learning on features through a BiLSTM network, including obtaining feature representations, performing forward LSTM calculations, performing backward LSTM calculations, and fusing the forward and backward outputs;

[0027] The local feature extraction unit is used to perform local feature extraction through a DPCNN network, including constructing a DPCNN network structure and a global pooling layer.

[0028] Furthermore, when the real-time data processing module extracts warning labels, it determines the warning labels by calculating the negative log similarity between the predicted probability distribution and the true labels.

[0029] Furthermore, the result output and model optimization module updates the model parameters according to the calculated probability difference to improve the prediction accuracy of the model.

[0030] Furthermore, when the single-stream model fuses features, it merges inputs of different modalities through ConCat, or performs feature fusion through element-wise addition operations.

[0031] A method for early warning detection of power work order visualization management includes the following steps:

[0032] Preprocess the historical work order data;

[0033] Train a first model based on the preprocessed historical work order data to extract features;

[0034] Construct a prediction model by fusing the features extracted by the first model. The prediction model uses a single-stream model to fuse text, audio, and structured features, and is optimized through Transformer layers and cross-entropy loss;

[0035] Input real-time data into the prediction model, extract warning labels, and calculate the probability difference;

[0036] Output the results and iteratively optimize the prediction model according to the real-time data.

[0037] Furthermore, the steps of preprocessing the historical work order data include data cleaning, word segmentation, and noise reduction processing;

[0038] In the step of training the first model, use the Word2Vec model to perform word vector training on the text data to obtain the feature representation of the text;

[0039] In the step of constructing the prediction model by fusing features, use the BiLSTM network to learn the temporal information of the features, and use the DPCNN network to extract local features.

[0040] Furthermore, in the step of extracting warning labels, determine the warning labels by calculating the negative logarithm similarity between the predicted probability distribution and the true labels;

[0041] In the step of iteratively optimizing the prediction model, update the model parameters according to the calculated probability difference to improve the prediction accuracy of the model;

[0042] When the single-stream model fuses features, it merges the inputs of different modalities through ConCat or other methods, or performs feature fusion through element-wise addition operations.

[0043] As can be seen from the above technical solutions, the present invention discloses a power work order visualization management warning detection system and method. Compared with the prior art, the first model training module of the present invention uses advanced algorithms such as Word2Vec to perform word vector training on the text data, accurately captures the text semantics, enables the model to deeply understand the key information in the work order text, such as fault description, processing process, etc., and provides strong support for accurate warning. The feature fusion module skillfully fuses the BiLSTM and DPCNN networks, fully excavates the temporal information and local features of the work order data, enables the model to not only grasp the context of the fault development but also pay attention to details, and comprehensively captures potential risks, significantly improving the accuracy of warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings.

[0045] Figure 1 It is a schematic structural diagram of a visual management early warning detection system for electric power work orders of the present invention.

[0046] Figure 2 It is a schematic flow diagram of a visual management early warning detection method for electric power work orders of the present invention. Specific embodiments

[0047] The following clearly and completely describes the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0048] Refer to the attached Figure 1-2 The embodiments of the present invention disclose a visual management early warning detection system and method for electric power work orders.

[0049] Embodiment 1 discloses the data preprocessing process of this application, mainly including data cleaning, word segmentation, and noise reduction;

[0050] Data cleaning includes removing error data, filling in missing data, and removing duplicate data;

[0051] Check the data format: Verify whether each field in the historical work order data conforms to the preset data format. For example, check whether the work order number is in the specified format of a combination of numbers and letters, and whether the fault time is in the standard date and time format, etc. Mark the data records that do not conform to the format as potential error data.

[0052] Data range verification: Conduct a range check on numerical fields, such as the equipment operation duration, fault handling duration, etc., to determine whether their values are within a reasonable range. If it exceeds the normal range, such as a negative or abnormally long fault handling duration, it is regarded as error data.

[0053] Logical consistency check: Analyze the logical relationship between fields to determine whether there are logical contradictions. For example, if the creation time of the work order is later than the fault handling completion time, there is a logical error in this record and it should be removed.

[0054] Numerical data imputation: For missing numerical fields, such as the operating parameters of equipment, select an appropriate imputation method according to its distribution. If the data is normally distributed, mean imputation can be used; if the data distribution is uneven, median imputation can be adopted. For example, if the voltage parameter of a certain device is missing, the mean value of the voltage parameters of this device in a similar time period can be used for imputation.

[0055] Categorical data imputation: For missing categorical fields, such as fault types, departments to which the handlers belong, etc., mode imputation can be used. That is, select the category with the highest frequency of occurrence in the dataset for imputation. If the fault type of a work order is missing and "equipment failure" is the type with the highest frequency of occurrence, then "equipment failure" is used to fill the missing value.

[0056] Model-based imputation: For complex missing situations, a prediction model can be constructed, such as using other complete fields to predict the values of missing fields. For example, based on information such as the fault description and equipment model of a work order, the missing fault handling result is predicted through a machine learning model.

[0057] Exact match deduplication: Perform exact matching on historical work order data to find exactly the same records. By comparing all fields or combinations of key fields, such as work order numbers, fault times, faulty equipment, etc., the exactly identical duplicate records are removed, and only one record is retained.

[0058] Fuzzy match deduplication: Considering the possible minor differences during data entry, such as individual typos and space differences in text descriptions, a fuzzy matching algorithm is used for deduplication. Set a certain similarity threshold. For example, when the similarity of two records exceeds 90%, they are determined to be duplicate records and removed.

[0059] Word segmentation includes selecting a word segmentation tool and dictionary and performing word segmentation operations;

[0060] Determine the word segmentation tool: According to the characteristics of power work order texts, select an appropriate word segmentation tool, such as jieba for Chinese texts and NLTK for English texts. These tools provide basic word segmentation functions and rich word segmentation modes.

[0061] Customize the power professional dictionary: Since power work orders contain a large number of professional terms, such as "transformer", "circuit breaker", "relay protection", etc., it is necessary to customize a power professional dictionary and load it into the word segmentation tool. This helps to improve the accuracy and professionalism of word segmentation and ensure that professional terms can be correctly segmented.

[0062] Word segmentation: Input the preprocessed text into the word segmentation tool, and cut the text into independent words according to the power professional dictionary and general language rules. For example, "10kV line single-phase grounding fault occurs" is cut into words such as "10kV, line, occurrence, single-phase, grounding, fault" and so on.

[0063] Filter stop words: In the word segmentation results, remove common stop words, such as "的", "了", "在", "和", etc. (in Chinese) or "the", "is", "at", "and", etc. (in English). Stop words usually contribute little to the semantics of the text. Removing them can reduce the amount of data and improve the efficiency of subsequent processing.

[0064] Noise reduction includes text noise reduction, audio noise reduction and structured data noise reduction;

[0065] Removing irrelevant information: In the work order text, there may be information irrelevant to the fault warning, such as some non-critical descriptions of the customer, notes of the staff, etc. Through methods such as keyword matching and rule filtering, these irrelevant information can be identified and removed, and the valid text content related to the fault can be retained. For example, irrelevant descriptions such as "the customer expressed great anxiety" and "the staff has contacted the customer" can be removed from the text.

[0066] Correcting typos and colloquial expressions: Using the pre-built typo dictionary and colloquial expression rule base for the power industry, typos in the text are corrected and colloquial expressions are converted into formal written language. For example, "电览" is corrected to "电线" and "那个什设备坏了" is converted to "该设备失失了".

[0067] Preprocess audio: Preprocess the audio data, such as removing silent segments in the audio, adjusting the audio sampling rate and bit depth, etc., to make it meet the requirements of subsequent noise reduction processing.

[0068] Application of noise reduction algorithm: Use professional audio noise reduction algorithms, such as spectral subtraction, Wiener filtering, etc., to suppress background noise in the audio. By analyzing the spectral characteristics of the audio, distinguishing between speech signals and noise signals, reducing the amplitude of noise, and improving the clarity of the audio, effective information in the audio can be extracted more accurately.

[0069] Data consistency check: For fields in structured data, such as equipment numbers and fault codes, consistency check is performed. Check whether the numbers of the same equipment in different work orders are consistent, and whether the fault codes corresponding to the same fault type are unified. If inconsistencies are found, corrections are made according to preset rules or through data tracing to ensure the consistency and accuracy of structured data.

[0070] Outlier handling: Identify outliers in structured data, such as obviously incorrect numerical values, data that does not conform to business logic, etc. For these outliers, they can be handled by deletion, correction, or replacement with other reasonable values. For example, if the recorded operating temperature of a device is an extremely unreasonable value, it can be deleted or replaced with the normal temperature value of the device during a similar time period.

[0071] Word vector training sub-step of the first model training module:

[0072] Build a vocabulary: Count all non-repeating words in the word segmentation results to build a vocabulary. At the same time, record the frequency of each word in the text data to provide basic data for subsequent word vector training. For example, the vocabulary contains words such as "transformer", "fault", "power outage", etc., and their corresponding frequency information.

[0073] Determine the word vector dimension: According to the characteristics of the power work order text data and the requirements of model training, select an appropriate word vector dimension. Generally speaking, the higher the dimension, the richer the semantic information that the word vector can represent, but the computational cost will also increase accordingly. For example, a 100-dimensional or 200-dimensional word vector can be selected.

[0074] Set the window size: The window size determines how many words in the vicinity of the current word will be considered as context words during training. The choice of window size needs to balance local semantic information and global semantic information. For example, a window size of 5 can be selected, that is, when training the word vector of the word "fault", the 5 words before and after it will be considered as context information.

[0075] Select a training algorithm: The Word2Vec model mainly has two training algorithms, namely CBOW (Continuous Bag of Words) and Skip-Gram. CBOW predicts the current word based on context words and is suitable for word vector training of common words; Skip-Gram predicts context words based on the current word and has better training results for rare words. Select an appropriate training algorithm according to the distribution of words in the power work order text. For example, if the text contains more professional terms and rare fault descriptions, the Skip-Gram algorithm can be selected.

[0076] Initialize word vectors: Randomly initialize a word vector for each word in the vocabulary as the starting point of training. These initial word vectors are usually vectors with the same dimension as the set word vector dimension and element values randomly distributed within a certain range.

[0077] Forward propagation calculation: Taking the CBOW algorithm as an example, for each word in the text, based on the word vectors of its context words, through the forward propagation process of the neural network, the predicted word vector of the current word is calculated. Specifically, the word vectors of the context words are weighted and summed or other operations are performed to obtain an intermediate vector, and then through a non-linear activation function, such as the softmax function, the intermediate vector is mapped to the predicted word vector of the current word.

[0078] Calculating the loss function: Compare the predicted word vector with the true word vector of the current word and calculate the value of the loss function. Commonly used loss functions include the cross-entropy loss function, etc. The value of the loss function reflects the difference between the predicted word vector and the true word vector. The greater the difference, the higher the value of the loss function.

[0079] Backpropagation update: According to the value of the loss function, calculate the gradient of each word vector through the backpropagation algorithm, and then update the word vector according to the gradient and the learning rate. The updated word vector can better represent the semantic information of the word, reducing the difference between the predicted word vector and the true word vector.

[0080] Iterative training: Repeat the above processes of forward propagation calculation, calculating the loss function, and backpropagation update, and perform multiple rounds of iterative training on the entire power work order text data. During the training process, the word vectors will be continuously updated and optimized until the value of the loss function converges to a small range or reaches the preset number of training rounds.

[0081] Evaluating the quality of word vectors: Through some evaluation metrics and methods, evaluate the quality of the trained word vectors. For example, the cosine similarity between words can be calculated to see if the word vectors of similar words are close in the high-dimensional space; also, through some simple text classification or clustering tasks, test the effect of the word vectors in practical applications and evaluate their ability to represent the characteristics of power work order texts.

[0082] Adjusting model parameters: According to the evaluation results of the word vectors, adjust and optimize the parameters of the model. If it is found that the quality of the word vectors is not ideal, parameters such as the word vector dimension, window size, training algorithm, etc. can be appropriately adjusted, and the word vectors are retrained until word vectors that meet the requirements are obtained. For example, if the evaluation finds that the word vectors do not accurately represent the words describing rare faults, the window size can be increased or the training algorithm can be adjusted to improve the quality of the word vectors.

[0083] Sub-steps of the feature fusion module:

[0084] The BiLSTM network learns the temporal information of the features

[0085] Input feature sequence: Combine the features extracted by the first model, such as the word vector features of text, the acoustic features of audio, and the numerical features of structured data, etc., into a feature sequence according to the time sequence or processing flow sequence of the work order, and input it into the BiLSTM network. For example, for a work order containing multiple processing steps, arrange the features corresponding to each step in the step sequence to form a feature sequence.

[0086] Forward LSTM calculation:

[0087] Initialize the hidden state: Initialize a zero vector as the initial hidden state for the forward LSTM, which is used to store the temporal information of the previous moment.

[0088] Calculate at each moment: Starting from the first feature of the feature sequence, sequentially input each feature and the hidden state of the previous moment into the forward LSTM cell. Inside each cell, through the calculations of the forget gate, input gate, and output gate, update the hidden state and output the hidden state of the current moment. The forget gate is used to determine how much information from the previous hidden state needs to be retained, the input gate is used to determine how much information from the current input feature needs to be written into the hidden state, and the output gate is used to determine how much information from the current hidden state needs to be output. For example, when processing the feature of the first processing step of the work order, the forget gate may forget some information of the previous state that is irrelevant to this step, the input gate writes the key feature information of this step into the hidden state, and the output gate outputs the hidden state containing the temporal information of this step.

[0089] Obtain the forward output: When all the features in the feature sequence have passed through the forward LSTM calculation, a forward hidden state sequence is obtained, which contains the temporal information from the start of the sequence to each moment.

[0090] Backward LSTM calculation:

[0091] Initialize the hidden state: Similar to the forward LSTM, initialize a zero vector as the initial hidden state for the backward LSTM.

[0092] Reverse calculation at each moment: Starting from the last feature of the feature sequence, reverse sequentially input each feature and the hidden state of the previous moment (in the reverse direction) into the backward LSTM cell, perform similar calculations of the forget gate, input gate, and output gate, update the hidden state, and output the hidden state of the current moment. For example, when processing the feature of the last processing step of the work order, the backward LSTM will update the hidden state according to the information of the subsequent steps (in the reverse direction) to capture the temporal information from the end of the sequence to each moment.

[0093] Obtain the reverse output: When all the features in the feature sequence have undergone reverse LSTM calculations, a reverse hidden state sequence is obtained, which contains the temporal information from the end of the sequence to each time step.

[0094] Fuse the forward and reverse outputs: The hidden states output by the forward LSTM and the reverse LSTM at each time step are fused, usually using methods such as concatenation, summation, or averaging. For example, the forward and reverse hidden states are concatenated into a higher-dimensional vector as the fused feature sequence. This fused feature sequence contains both the temporal information from the start of the sequence to each time step and the temporal information from the end of the sequence to each time step, and can capture the temporal dependencies in the feature sequence more comprehensively.

[0095] The DPCNN network performs local feature extraction

[0096] Construct the DPCNN network structure:

[0097] One-dimensional convolutional layer: Apply a one-dimensional convolutional layer to the fused feature sequence to extract local features. Set appropriate convolutional kernel sizes and numbers. The convolutional kernel slides over the feature sequence and performs operations such as weighted summation on the features in each local region to generate new feature representations. For example, a convolutional kernel size of 3 means that the features of 3 consecutive time steps are considered for convolution each time, and it can capture the feature correlations between local time steps.

[0098] Activation function: Add an activation function such as ReLU after the one-dimensional convolutional layer to introduce non-linearity, enabling the network to learn more complex feature patterns. The activation function performs a non-linear transformation on the convolved features, enhancing the expressive power of the features. For example, the ReLU function can set the negative-valued features after convolution to 0 and retain the positive-valued features, making the features sparser and highlighting important features.

[0099] Pooling layer: Add a pooling layer after the one-dimensional convolutional layer and the activation function to reduce the feature dimension and extract the main feature information. The pooling layer performs downsampling on the features in the local region through operations such as max pooling or average pooling, retains the key features, and reduces the computational amount and model complexity. For example, max pooling selects the maximum feature value in the local region as the representative to highlight the significant features.

[0100] Repeat convolution and pooling: Repeat the operations of the above one-dimensional convolutional layer, activation function, and pooling layer multiple times to form multiple convolutional blocks. Each convolutional block further abstracts and extracts the features, enabling the network to learn higher-level local features. For example, the first convolutional block may extract simple feature correlations at local time steps, and the second convolutional block may extract more complex feature combination patterns.

[0101] Global pooling layer:

[0102] Apply global pooling: After extracting local features through multiple convolutional blocks, apply a global pooling layer to perform global pooling on the final feature map. The global pooling layer pools the entire feature map and reduces the feature map to a feature vector of a fixed length. For example, global max pooling selects the maximum feature values at all positions in the feature map to form the feature vector, and global average pooling calculates the average of all feature values in the feature map to form the feature vector.

[0103] Obtain the final features: The resulting feature vector of a fixed length is the local feature representation extracted by the DPCNN network. It can capture the local feature patterns and key information in the fused feature sequence, providing effective feature inputs for the construction of further prediction models.

[0104] Sub-steps for the real-time data processing module to extract warning labels:

[0105] Real-time data input and feature extraction

[0106] Receive real-time data: The real-time data processing module first receives the input real-time power work order data, which may include text descriptions, audio recordings, and relevant structured information, etc.

[0107] Feature extraction: Use the pre-trained first model and the feature fusion module to extract features from the real-time data. Convert the text in the real-time data into text features through operations such as word segmentation and word vector conversion; extract acoustic features from the audio data; at the same time, perform corresponding numerical processing on the structured data. Then, input these multi-modal features into the BiLSTM network and the DPCNN network to extract a comprehensive feature representation that fuses temporal information and local features.

[0108] Input the prediction model: Input the extracted comprehensive feature representation into the constructed prediction model. The prediction model can be a deep learning-based classification model, such as a fully connected neural network, a convolutional neural network, etc. Its purpose is to predict the corresponding warning labels based on the input features.

[0109] Calculate the predicted probability distribution: The prediction model calculates through forward propagation and outputs the predicted probabilities of each possible warning label. These probability values form a probability distribution, indicating the prediction confidence of the model for each warning label. For example, for a classification problem with multiple warning labels such as "equipment failure", "line aging", and "human error", the probability distribution output by the model may be [0.7, 0.2, 0.1], indicating that the model predicts the probability of "equipment failure" is the highest, 70%, the probability of "line aging" is 20%, and the probability of "human error" is 10%.

[0110] Calculate the negative log similarity

[0111] Determine the true label probability distribution: In practical applications, there is usually a true label or a reference label determined based on expert knowledge, historical data, etc. Convert this true label or reference label into a probability distribution. If the label is certain, the corresponding probability is 1, and the probabilities of other labels are 0. For example, if the true label is "equipment failure", the true label probability distribution is [1, 0, 0].

[0112] Calculate the negative logarithm probability: For each probability value in the predicted probability distribution, calculate its negative logarithm. The negative logarithm operation can convert the probability value into a positive value, and the smaller the probability, the larger the negative logarithm. For example, for the predicted probability 0.7, its negative logarithm is -log(0.7) ≈ 0.357; for the predicted probability 0.2, its negative logarithm is -log(0.2) ≈ 1.609.

[0113] Calculate the similarity: Compare the negative logarithm of the predicted probability distribution with the true label probability distribution and calculate the similarity between them. Various similarity calculation methods can be used, such as Euclidean distance, Manhattan distance, cosine similarity, etc. Taking cosine similarity as an example, calculate the cosine value of the angle between the negative logarithm vector of the predicted probability distribution and the true label probability distribution vector. The closer the cosine value is to 1, the more similar the two distributions are.

[0114] Determine the most likely warning label

[0115] Evaluate the similarity: Based on the calculated negative logarithm similarity, evaluate the matching degree between the predicted probability distribution and the true label probability distribution. If the similarity is higher than a preset threshold, it indicates that the predicted probability distribution is relatively close to the true label, and the prediction of the model is relatively reliable.

[0116] Select the most likely warning label: In the predicted probability distribution, select the label with the largest probability value and the corresponding negative logarithm similarity meeting the requirements as the most likely warning label. For example, in the above probability distribution [0.7, 0.2, 0.1], if after calculating the negative logarithm similarity, it is found that the similarity of the "equipment failure" label meets the requirements, then determine "equipment failure" as the most likely warning label.

[0117] Output the warning result: Output the determined most likely warning label, providing a basis for the processing and warning of power work orders. Relevant personnel can take corresponding measures in a timely manner according to this warning label, such as arranging maintenance personnel, preparing spare parts, etc., to deal with possible power failure problems.

[0118] Sub - steps of the result output and model optimization module

[0119] Calculate the probability difference

[0120] Determine the key probability value:

[0121] Predicted Probability: Obtain the predicted probability of the most likely warning label from the real-time data processing module, which is the probability value that the model predicts this label as the correct result. For example, the probability that the model predicts the warning label of "equipment failure" is 0.8.

[0122] True Probability (if available): If the probability corresponding to the true label can be obtained in the actual application scenario (usually can be obtained during the training phase, but may not be directly obtainable during the actual prediction phase), then record this true probability. During the training phase, the true probability is 1 (for the correct label) or 0 (for the wrong label).

[0123] Second-Highest Probability: Record the second-highest probability value in the predicted probability distribution, which is used to measure the prediction confidence of the model for other possible results. For example, the second-highest predicted probability is 0.15, corresponding to the label of "line aging".

[0124] Calculate the Probability Difference:

[0125] Predicted Probability Difference: Calculate the difference between the predicted probability of the most likely warning label and the second-highest probability. This difference reflects the degree of differentiation of the model between the most likely result and other results. The larger the difference, the more confident the model is in predicting the most likely result. For example, 0.8 (probability of the most likely result) - 0.15 (second-highest probability) = 0.65.

[0126] Difference from True Probability: If there is a true probability, calculate the difference between the predicted probability of the most likely warning label and the true probability. During the training phase, this can intuitively reflect the accuracy of the model prediction. For example, 0.8 (predicted probability) - 1 (true probability) = -0.2.

[0127] Evaluate Model Performance

[0128] Set Threshold Criteria:

[0129] Predicted Probability Difference Threshold: According to business requirements and model performance goals, set a threshold for the predicted probability difference. When the predicted probability difference is greater than this threshold, it is considered that the model's prediction of the most likely warning label is relatively reliable; otherwise, it is considered that the prediction result is not certain enough and the model needs to be further optimized. For example, set the threshold to 0.5.

[0130] Accuracy Threshold: If in the training phase or a supervised prediction scenario, set a threshold for the model accuracy. The accuracy can be calculated by comparing the model prediction results with the true results. When the model accuracy is lower than this threshold, trigger the model optimization process.

[0131] Analysis of Evaluation Results:

[0132] Prediction Probability Difference Evaluation: Compare the calculated prediction probability difference with the set threshold. If the difference is greater than the threshold, output the result and do not optimize the model temporarily; if the difference is less than or equal to the threshold, enter the model optimization process.

[0133] Accuracy Evaluation: Calculate the accuracy of the model on a certain dataset and compare it with the accuracy threshold. If the accuracy is lower than the threshold, it indicates that the overall performance of the model is poor, and it is necessary to optimize the model parameters to improve the prediction accuracy.

[0134] Model Parameter Update

[0135] Select Optimization Algorithm:

[0136] Gradient Descent Algorithms: Commonly used ones include Stochastic Gradient Descent (SGD), Batch Gradient Descent (BGD), Mini-Batch Gradient Descent (MBGD), etc. These algorithms update the model parameters by calculating the gradients of the loss function with respect to the model parameters, reducing the value of the loss function, thereby optimizing the model performance.

[0137] Momentum Optimization Algorithm: Such as Nesterov Accelerated Gradient (NAG), which introduces a momentum term on the basis of gradient descent, can accelerate the model convergence and avoid local optima to a certain extent.

[0138] Adaptive Learning Rate Algorithms: Such as Adagrad, RMSprop, Adam, etc. These algorithms can automatically adjust the learning rate according to the historical gradient information of the model parameters, enabling the model to have different learning speeds in different parameter dimensions, improving the optimization efficiency and effect.

[0139] Calculate Gradient and Update Parameters:

[0140] Loss Function Calculation: Calculate the value of the loss function according to the deviation between the model prediction result and the true result (in the training stage) or the business objective. Commonly used loss functions include mean squared error loss, cross-entropy loss, etc. The value of the loss function reflects the degree of inaccuracy of the model prediction.

[0141] Gradient Calculation: Use the backpropagation algorithm to calculate the gradients of the loss function with respect to each model parameter. The gradient value represents the change direction and rate of the loss function in the parameter space, providing a basis for parameter update.

[0142] Parameter Update: Update the model parameters according to the selected optimization algorithm and the calculated gradient value. For example, in SGD, the parameter update formula is: parameter = parameter - learning rate × gradient. By continuously iteratively updating the parameters, the value of the loss function is gradually reduced, and the model prediction accuracy is improved.

[0143] Model Optimization and Verification

[0144] Iterative Optimization:

[0145] Multiple iterations: Repeat the above process of calculating probability difference, evaluating model performance, and updating model parameters multiple times, using new parameters for prediction and evaluation each time. As the number of iterations increases, the model parameters are continuously optimized and the prediction accuracy gradually improves.

[0146] Early stopping strategy: To avoid overfitting of the model, set an early stopping strategy during the iteration process. When the performance of the model on the validation set no longer improves, or improves very slowly, stop the iteration and retain the current optimal model parameters. For example, if the accuracy of the model on the validation set does not improve significantly after 10 consecutive iterations, stop training.

[0147] Model Validation and Evaluation:

[0148] Validation set evaluation: During the model optimization process, an independent validation set is used to evaluate the model. The validation set should be different from the training set and can truly reflect the performance of the model on unknown data. Calculate various indicators of the model on the validation set, such as accuracy, recall, F1 value, etc., to comprehensively evaluate the model performance.

[0149] Cross-validation: If the amount of data allows, the cross-validation method can be used to evaluate the model more rigorously. Divide the data set into multiple subsets, take one of the subsets as the validation set and the other subsets as the training set in turn, train and validate the model multiple times, and take the average performance index as the final evaluation result of the model to improve the reliability and stability of the model evaluation.

[0150] Sub-steps of single-stream model feature fusion

[0151] Feature extraction and preparation

[0152] Text feature extraction:

[0153] Word segmentation and word embedding: Perform word segmentation on the power work order text data and cut the text into word sequences. Then, use a pre-trained word vector model (such as Word2Vec, GloVe) or a trained text model to convert the word sequence into a word embedding vector sequence. Each word embedding vector represents the semantic features of a word.

[0154] Text sequence encoding: Encode the word embedding vector sequence through a recurrent neural network (such as LSTM, GRU) or a Transformer encoder to obtain the contextual representation of the text sequence. For example, using a BiLSTM network, you can consider both the forward and reverse context information of the text sequence and obtain a context-aware representation of each word.

[0155] Text feature pooling: Perform pooling operations on the encoded text sequence, such as max pooling, average pooling, etc., to compress the sequence features into a fixed-length text feature vector. This vector can capture the key information in the text and serve as the feature representation of the text modality.

[0156] Audio feature extraction:

[0157] Acoustic feature calculation: For audio data related to power work orders (such as on-site fault recordings, customer call recordings, etc.), calculate acoustic features such as Mel Frequency Cepstral Coefficients (MFCC), spectrograms, etc. These features can reflect the spectral characteristics, time-domain characteristics, etc. of the audio signal.

[0158] Audio sequence encoding: Input the acoustic feature sequence into an audio encoding network, such as a Convolutional Neural Network (CNN) or a Recurrent Neural Network, to extract the high-level features of the audio. For example, through CNN, local patterns and structural information in the audio signal can be learned.

[0159] Audio feature pooling: Perform pooling operations on the encoded audio feature sequence to obtain a fixed-length audio feature vector, which serves as the feature representation of the audio modality.

[0160] Structured feature processing:

[0161] Numerical feature normalization: For numerical structured data in power work orders (such as equipment operation parameters, fault handling duration, etc.), perform normalization processing to scale it to a certain range (such as between 0 and 1) to eliminate the dimensional differences between different numerical features and facilitate model learning.

[0162] Categorical feature encoding: For categorical structured data (such as fault types, equipment models, etc.), perform one-hot encoding or embedding encoding to convert it into a numerical feature vector. For example, convert fault types such as "equipment failure" and "line aging" into vector forms such as [1,0] and [0,1] through one-hot encoding.

[0163] Structured feature integration: Perform operations such as concatenation or weighted summation on the processed numerical features and categorical features to obtain a comprehensive representation of the structured features.

[0164] Feature fusion method selection

[0165] ConCat method:

[0166] Feature Dimension Alignment: Before performing feature concatenation, ensure that the dimensions of different modality features are consistent or compatible. If the feature dimensions are different, the features can be mapped to the same dimensional space through linear transformation (such as a fully connected layer). For example, transform the text feature dimension from 256 dimensions to 128 dimensions, keep the audio feature dimension unchanged at 128 dimensions, and transform the structured feature dimension from 64 dimensions to 128 dimensions.

[0167] Feature Concatenation: Concatenate the dimension-aligned different modality features along a specific dimension (such as the last dimension) to form a higher-dimensional fused feature vector. For example, concatenate the text feature, audio feature, and structured feature as the rows or columns of a matrix respectively to obtain a feature representation that combines multi-modal information.

[0168] Element-wise Addition Method:

[0169] Feature Normalization: Before performing element-wise addition, normalize different modality features so that they have similar numerical ranges and distribution characteristics. This helps to avoid certain feature information being masked due to large numerical differences in the addition operation of different modality features. For example, normalize the text feature, audio feature, and structured feature to the range of [-1, 1] or [0, 1].

[0170] Feature Weighting: Assign different weights to each modality feature according to its importance or relevance. The weights can be pre-set or learned. For example, if the text feature is more critical in power work order warning, a larger weight can be assigned to the text feature, such as 0.5, 0.3 to the audio feature, and 0.2 to the structured feature.

[0171] Element-wise Addition Fusion: Perform element-wise addition operation on the normalized and weighted different modality features, that is, add the corresponding elements to obtain the fused feature vector. This vector combines the information of each modality feature and can represent the comprehensive features of power work orders more comprehensively.

[0172] Subsequent Processing of Fused Features

[0173] Feature Dimensionality Reduction (Optional):

[0174] Principal Component Analysis (PCA): If the dimension of the fused feature is high, dimensionality reduction techniques such as PCA can be used to reduce the dimension of the feature. PCA projects the feature onto the principal component direction through linear transformation, retains the main variance information, removes redundant information, reduces the feature dimension, and minimizes information loss as much as possible. For example, reduce the fused feature from 512 dimensions to 128 dimensions for subsequent model processing and learning.

[0175] Autoencoder: Use autoencoder neural network to reduce the dimension of fused features. The autoencoder encodes high-dimensional features into low-dimensional feature representations through the encoder part, and then tries to reconstruct the original features through the decoder part. During the training process, the low-dimensional features learned by the autoencoder can capture the main information of the original features and achieve feature dimensionality reduction.

[0176] Features are input to the prediction model:

[0177] Build a prediction model: According to the requirements of the power work order warning task, build a suitable prediction model, such as a fully connected neural network, a convolutional neural network, a recurrent neural network, etc. The input layer dimension of the prediction model should match the dimension of the fusion feature.

[0178] Model training and optimization: Input the fusion features into the prediction model and train the model through supervised learning. Use the labeled power work order data as training samples, take the warning label as the true label, and calculate the loss function (such as cross entropy loss) between the model prediction result and the true label. According to the loss function value, use the back propagation algorithm and optimization algorithm (such as Adam) to update and optimize the model parameters so that the model can accurately predict the warning label of the power work order.

[0179] Model evaluation and application: Evaluate the trained prediction model on the validation set or test set, and calculate the model's performance indicators, such as accuracy, recall, F1 value, etc. If the model performance meets the requirements, the model can be applied to the actual power work order warning task, extract fusion features based on the real-time input work order data, and output warning labels through the prediction model to provide support for fault warning and processing of the power system.

[0180] The electric power work order visual management early warning detection system and method designed according to the present invention effectively improves the data quality through the refined operation of the data preprocessing module, lays a solid foundation for subsequent analysis, and enables the model to learn based on more accurate, complete and consistent data, thereby improving the reliability of prediction. Secondly, the first model training module uses advanced algorithms such as Word2Vec to train word vectors on text data, accurately captures the semantics of the text, and enables the model to deeply understand the key information in the work order text, such as fault description, processing process, etc., providing strong support for accurate early warning.

[0181] Furthermore, the feature fusion module cleverly integrates BiLSTM and DPCNN networks to fully explore the time series information and local features of the work order data, so that the model can not only grasp the context of the fault development, but also pay attention to details, capture potential risks in all directions, and significantly improve the accuracy of early warning. In addition, the real-time data processing module combines negative logarithmic similarity calculation to quickly and accurately extract the most likely early warning labels from real-time data, buying valuable time for the power system's immediate response and fault handling.

[0182] Finally, the result output and the model optimization module dynamically adjust the model parameters based on the probability difference to achieve the self - evolution and optimization of the model, enabling it to continuously adapt to the changes in power work order data, always maintain good prediction performance, provide long - term and reliable early warning guarantees for the stable operation of the power system, assist power enterprises in reducing costs and increasing efficiency, and improving the operation and maintenance management level.

[0183] In this specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other.

[0184] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A visual management warning detection system for power work orders, characterized in that Including: A data preprocessing module for preprocessing historical work order data; A first model training module for training a first model based on the preprocessed historical work order data to extract features; A feature fusion module for fusing the features extracted by the first model to construct a prediction model; A real-time data processing module for inputting real-time data into the prediction model, extracting warning labels, and calculating the probability difference; A result output and model optimization module for outputting results and iteratively optimizing the prediction model according to real-time data; Wherein, the prediction model uses a single-stream model to fuse text, audio, and structured features, and is optimized through Transformer layers and cross-entropy loss.

2. The visual management warning detection system for electric power work orders according to claim 1, wherein The data preprocessing module includes a data cleaning unit, a word segmentation unit, and a noise reduction unit; The data cleaning unit is used to remove incorrect data, duplicate data, and fill in missing data; The word segmentation unit is used to select a word segmentation tool and dictionary and perform word segmentation operations; The noise reduction unit is used for text noise reduction, audio noise reduction, and structured data noise reduction.

3. The visual management warning detection system for power work orders according to claim 1, wherein The first model training module uses a Word2Vec model to perform word vector training on text data to obtain the feature representation of the text, including a model parameter setting unit, a word vector training unit, and an evaluation and optimization unit; The model parameter setting unit is used to determine the word vector dimension, set the window size, and select a training algorithm; The word vector training unit is used to initialize the word vector, perform forward propagation calculation, calculate the loss function, perform backpropagation update, and iteratively train; The evaluation and optimization unit is used to evaluate the quality of the word vector and adjust the model parameters.

4. The visual management warning detection system for power work orders according to claim 1, wherein, The feature fusion module includes a temporal information learning unit and a local feature extraction unit; The temporal information learning unit is used to perform temporal information learning on features through a BiLSTM network, including obtaining feature representations, performing forward LSTM calculations, performing backward LSTM calculations, and fusing the forward and backward outputs; The local feature extraction unit is used to perform local feature extraction through a DPCNN network, including constructing a DPCNN network structure and a global pooling layer.

5. The visual management early warning detection system for power work orders according to claim 1, wherein, When the real-time data processing module extracts warning labels, it determines the warning labels by calculating the negative log similarity between the predicted probability distribution and the true labels.

6. The visual management early warning detection system for power work orders according to claim 1, characterized in that, The result output and model optimization module updates the model parameters according to the calculated probability difference.

7. The early warning detection system for visual management of power work orders according to claim 1, characterized in that, When the single-stream model fuses features, it merges the inputs of different modalities through ConCat or performs feature fusion through element-wise addition operations.

8. A visual management warning detection method for power work orders, applicable to any one of the visual management warning detection systems for power work orders described in claims 1-7, characterized in that, Including the following steps: Preprocess the historical work order data; Train a first model based on the preprocessed historical work order data to extract features; Fuse the features extracted by the first model to construct a prediction model, where the prediction model uses a single-stream model to fuse text, audio, and structured features, and is optimized through Transformer layers and cross-entropy loss; Input real-time data into the prediction model, extract warning labels, and calculate the probability difference; Output the results and iteratively optimize the prediction model according to real-time data.

9. A power work order visualization management warning detection method according to claim 8, characterized in that The steps for preprocessing the historical work order data include data cleaning, word segmentation, and noise reduction processing; In the steps of training the first model, the Word2Vec model is used to perform word vector training on the text data to obtain the feature representation of the text; In the steps of constructing the prediction model by fusing features, the BiLSTM network is used to learn the temporal information of the features, and the DPCNN network is used to extract local features.

10. A power work order visualization management early warning detection method according to claim 8, wherein, In the steps of extracting the early warning labels, the early warning labels are determined by calculating the negative logarithmic similarity between the predicted probability distribution and the true labels; In the steps of iteratively optimizing the prediction model, the model parameters are updated according to the calculated probability difference; When the single-stream model fuses features, different modalities of inputs are merged by ConCat, or feature fusion is performed through element-wise addition operations.

Images