An artificial intelligence-based digital system work order auditing method and system

Abstract

The present invention relates to the field of intelligent operation and maintenance management technology, specifically a digital system work order review and method based on artificial intelligence, which obtains fields to construct information sets, extracts features to form vector tables, classifies and identifies to generate review results and outputs them in a structured manner. The present invention improves the integrity and consistency of information expression by standardizing the text description and image content in the work order and integrating semantic and image state features. The combination of semantic analysis and image recognition improves the ability to judge the accuracy of the content. Priority level information is integrated into the construction of semantic and image feature vectors to achieve the fusion expression of multi-dimensional information. The matching detection of graphic and text content is enhanced with the help of vector similarity calculation to ensure the accuracy of the graphic and text consistency assessment in the work order review. The support vector machine model performs status and risk level identification on this basis, making the review label and risk judgment more accurate.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent operation and maintenance management, and in particular to a digital system work order auditing method based on artificial intelligence. BACKGROUND

[0002] As an important carrier of renewable energy, the operation and maintenance (O&M) of photovoltaic power stations is significantly improved in complexity with the increase of installed capacity and the number of stations. The work order system undertakes the flow management of tasks such as fault repair, preventive maintenance, inspection, and upgrading, and work order auditing is a key link that is essential to ensuring the accuracy, integrity, and standardization of work order information. At present, most photovoltaic operation and maintenance units still use manual auditing methods, which have low auditing efficiency, subjective judgment bias, high error rate, rising labor costs, and difficulty in systematizing knowledge. Although some units have introduced rule-based automated auditing mechanisms, they rely on static "IF-THEN" rules, lack flexibility and generalization ability, and are difficult to adapt to the characteristics of photovoltaic operation and maintenance scenarios, such as varying fault modes and complex heterogeneous data, thus having obvious shortcomings in handling real-time and scalability requirements. Therefore, a work order auditing scheme with high intelligence and sustainable learning ability needs to be developed to meet the urgent needs of efficient, low-error, and intelligent auditing in new photovoltaic operation and maintenance management. SUMMARY

[0003] The application aims to solve the problems in the prior art and provides a digital system work order auditing method based on artificial intelligence.

[0004] To achieve the above-mentioned purpose, the application adopts the following technical scheme: a digital system work order auditing method based on artificial intelligence, comprising the following steps:

[0005] S1: Obtain the fault description field and image field of the photovoltaic power station operation and maintenance work order, and perform standardization processing to construct a photovoltaic power station work order information set;

[0006] S2: Extract the semantic features of the fault description field and the state features of the image field in the photovoltaic power station work order information set, and construct an operation and maintenance work order feature vector table;

[0007] S3: Input the operation and maintenance work order feature vector table into a support vector machine classification model, identify the auditing state of the operation and maintenance work order, and divide the corresponding risk level to obtain a photovoltaic power station auditing result;

[0008] S4: Integrate the auditing state, risk level, and corresponding field description in the photovoltaic power station auditing result to generate a structured auditing conclusion record.

[0009] The application improves that the photovoltaic power station work order information set comprises a standardized fault description field, a standardized image field and a device code field, the operation and maintenance work order feature vector table comprises a semantic embedding vector, an image state feature vector and a normalized priority field, the photovoltaic power station audit result is specifically an audit label value, a risk level identifier and a text-image consistency score, and the structured audit conclusion record comprises an audit state text, a risk level description and a field description.

[0010] The application improves that the specific steps of S1 are as follows:

[0011] S101: Obtain the work order fault description field, the device number field, the priority level field and the field image recorded by the photovoltaic power station digital platform, uniformly process the format of each type of field, and generate a field preliminary cleaning result;

[0012] S102: Based on the device number field in the field preliminary cleaning result, perform format matching and naming adjustment according to the component standard coding rule, and construct a photovoltaic power station work order information set.

[0013] The application improves that the specific steps of S2 are as follows:

[0014] S201: Obtain the fault description field in the photovoltaic power station work order information set, identify the component name, the alarm type and the operation behavior word group in the text through semantic analysis, embed the identified word group into a trainable vector representation through a bidirectional encoder representation model (BERT), and generate a semantic embedding result;

[0015] S202: Obtain the image field in the photovoltaic power station work order information set, identify the structure edge, color block abnormality and damage texture area in the image, extract the state feature vector of the corresponding area through a residual neural network (ResNet50) model, and obtain an image state feature extraction result;

[0016] S203: Based on the semantic embedding result and the image state feature extraction result, perform normalization processing and feature splicing according to the priority level field in the photovoltaic power station work order information set, integrate the vector set of the text feature, the image feature and the level information, and construct an operation and maintenance work order feature vector table.

[0017] The application improves that the specific steps of S3 are as follows:

[0018] S301: Obtain the corresponding text vector and image vector in the operation and maintenance work order feature vector table, calculate the cosine similarity between the vectors based on the feature space consistency between the semantic embeddings, identify the text-image matching degree of the work order content, and generate a text-image similarity calculation result;

[0019] S302: input the text similarity calculation result and the operation and maintenance work order feature vector table into a support vector machine classification model to perform classification identification of the audit state label, and obtain an audit label identification result;

[0020] S303: based on the audit label identification result, a priority field and a violation feature content in the operation and maintenance work order feature vector table, a scoring input set is constructed, a scoring interval is mapped according to a preset risk division rule, an audit state and a risk level are combined, and a photovoltaic power station audit result is obtained.

[0021] The application improves that the specific steps of S4 are as follows:

[0022] S401: obtain the audit state and risk level identification in the photovoltaic power station audit result, identify the corresponding text label and level information, analyze the audit state text and extract the work order classification keyword, and generate an audit state analysis result;

[0023] S402: based on the audit state analysis result and the operation and maintenance work order feature vector table, keyword extraction is performed on the field items of the specified ranking, and field description text is constructed by combining the fault description field and the labeled content in the image recognition area, and a field description generation result is obtained;

[0024] S403: based on the field description generation result and the audit state analysis result, the audit state text, the risk level description and the field content are arranged in sequence, and a structured audit conclusion record is generated.

[0025] The application improves that it also includes S5: based on the structured audit conclusion record, the label consistency and field expression of the audit conclusion content of the operation and maintenance work order are optimized, and an optimized audit conclusion content is obtained.

[0026] The specific indicators of the optimized audit conclusion content include a label style unification item, a semantic expression compression item and a field order adjustment item.

[0027] The application improves that the specific steps of S5 are as follows:

[0028] S501: obtain the audit state text, the risk level description and the field content in the structured audit conclusion record, identify the repeated semantic segments and the redundant field description, and generate an expression compression result;

[0029] S502: based on the expression compression result, perform style consistency processing on the keywords and label content in the reserved field, map and adjust the description style and semantic boundary according to the label standard, standardize the word range and expression form, obtain a label uniform processing result;

[0030] S503: Based on the label unified processing result, reorganize the field position according to the display order of the audit state, risk level and field description, and obtain the optimized audit conclusion content.

[0031] Compared with the prior art, the application has the advantages and positive effects that:

[0032] In the application, by standardizing the text description and image content in the work order and fusing semantic and image state features, the integrity and consistency of information expression are improved, the combination of semantic analysis and image recognition improves the judgment ability of content accuracy, the priority level information is fused in the construction of semantic and image feature vectors, the fusion expression of multi-dimensional information is realized, the matching detection of graphic and text content is strengthened by vector similarity calculation, the accuracy of graphic and text consistency evaluation in work order auditing is ensured, the support vector machine model identifies the state and risk level on this basis, so that the audit label and risk judgment are more accurate, the risk score is constructed by fusing the weight of multiple key fields to realize dynamic regulation, so that the risk assessment result is more detailed and adjustable, finally the audit state, risk level and field description are generated into structured conclusion, and the label style and expression order are unified, so that the standardization and readability of the audit conclusion are ensured, the system is enhanced in aspects of multi-source heterogeneous data fusion, audit efficiency improvement, error rate reduction and audit conclusion standardized expression, and the comprehensive needs of photovoltaic operation and maintenance work order auditing in efficiency, accuracy and intelligent degree are met. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The method flowchart of the application is shown in the figure;

[0034] Figure 2 The detailed flowchart of step S1 of the application is shown in the figure;

[0035] Figure 3 The detailed flowchart of step S2 of the application is shown in the figure;

[0036] Figure 4 The detailed flowchart of step S3 of the application is shown in the figure;

[0037] Figure 5 The detailed flowchart of step S4 of the application is shown in the figure;

[0038] Figure 6 The detailed flowchart of step S5 of the application is shown in the figure. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical scheme and advantages of the application clearer, the application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0040] In the description of the present application, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, which are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0041] Referring to Figure 1 The present application provides a technical solution: a digital system work order auditing method based on artificial intelligence, comprising the following steps:

[0042] S1: obtaining the fault description field and image field of the photovoltaic power station operation and maintenance work order, and performing standardization processing to construct a photovoltaic power station work order information set;

[0043] S2: extracting the semantic features of the fault description field and the state features of the image field in the photovoltaic power station work order information set, and constructing an operation and maintenance work order feature vector table;

[0044] S3: inputting the operation and maintenance work order feature vector table into a support vector machine classification model, identifying the audit state of the operation and maintenance work order and dividing the corresponding risk level, and obtaining the photovoltaic power station audit result;

[0045] S4: integrating the audit state, risk level and corresponding field description in the photovoltaic power station audit result to generate a structured audit conclusion record;

[0046] The photovoltaic power station work order information set includes standardized fault description field, standardized image field, equipment code field, the operation and maintenance work order feature vector table includes semantic embedding vector, image state feature vector, normalized priority field, the photovoltaic power station audit result is specifically audit label value, risk level identification, image-text consistency score, and the structured audit conclusion record includes audit state text, risk level description, field description.

[0047] Referring to Figure 2 The specific steps of S1 are as follows:

[0048] S101: obtaining the work order fault description field, equipment number field, priority level field and field image field recorded by the photovoltaic power station digital platform, uniformly processing the format of each type of field, and generating a field preliminary cleaning result;

[0049] Obtain the work order fault description field, equipment number field, priority level field and on-site image field recorded in the digital platform of the photovoltaic power station. In the process of cleaning the fields, first export the original record table containing the above fields from the platform database, use the Pandas library in Python to read the CSV format data, and replace and remove the empty values ​​and abnormal symbols in the fields. For example, if special characters or extra line breaks are found in the fault description field, they need to be removed uniformly through regular expressions. Then, the "equipment-001", "equipment 001" and "EQP_001" in the equipment number field due to different work order entries are uniformly processed. They can be converted into the form of "EQP001" using string replacement and standardization rules. The standardization of the priority level field needs to be combined with the level mapping relationship set by the platform. For example, "urgent", "general" and "low" are mapped to the values ​​3, 2 and 1 in turn for digital processing. The on-site image field format cleaning includes the route The paths are unified, and the image file format is checked and corrected. If there are inconsistent expressions such as "JPEG" and ".jpg", they are all converted to the standard ".jpg" format. At the same time, it is checked whether the file actually exists in the corresponding storage path. For example, Python's os.path.exists(path) is used to verify the validity of the image path. The default identifier "image_missing.jpg" is filled in for the missing image record field. The field set obtained through the above processing operation is the preliminary cleaning result. This process can be applied to any photovoltaic power station digital platform built on MySQL or Oracle database. For example, the field result after cleaning of the record of a power station work order number "WO2024061203" is: the fault description is "inverter communication interruption", the device number is "INV023", the priority level is 2, and the image field is " / images / WO2024061203_INV023.jpg".

[0050] S102: Based on the equipment number field in the preliminary field cleaning result, format matching and naming adjustment are performed according to the component standard coding rules to construct a photovoltaic power station work order information set;

[0051] Based on the equipment number field in the preliminary cleaning result of the field, in the process of constructing the photovoltaic power station work order information set, first of all, a set of component standard coding rules need to be established, for example, the standard rule is defined as "device type prefix + 3-digit serial number", that is, "INV023" in "INV" is the prefix of the inverter, and "023" is the equipment number. When performing format matching, the Python regular expression re.match(r"[A-Z]{3}\d{3}", device_id) is used to screen out the device numbers that meet the rules. For those that do not match, such as "INVERTER23", manual judgment or algorithm error correction matching to the standard format "INV023" is required. This process can be combined with the fuzzy matching technology Levenshtein distance, and the distance threshold is set to 2. If the edit distance between "INVERTER23" and "INV023" is 5, they do not match. Otherwise, the prefix is extracted and the serial number is reconstructed to the standard number by zero padding. Further combined with other field data in the field cleaning result, the structured work order information set is constructed using data structures such as dictionaries dict. Each work order record is set in the form of key-value pair, where the key is the work order number "WO2024061203", and the value is a nested dictionary containing the fields "equipment number", "fault description", "priority" and "image path". For example: {"WO2024061203":{"equipment number":"INV023","fault description":"inverter communication interruption","priority":2,"image path":" / images / WO2024061203_INV023.jpg"}}, the finally constructed work order information set can be imported into the database or platform interface call and used for subsequent data annotation and scheduling process.

[0052] Please refer to Figure 3 , the specific steps of S2 are as follows:

[0053] S201: Obtain the fault description field in the photovoltaic power station work order information set, identify the component name, alarm type and operation behavior phrase in the text through semantic parsing, and embed the identified phrases into trainable vector representation through the bidirectional encoder representation model BERT to generate semantic embedding results;

[0054] Get the fault description field in the photovoltaic power station work order information set. First, the natural language text data in the field is read in batches and stored in the text processing module. Taking the instance work order "WO2024061203" as an example, the fault description is "Component E23 appears voltage anomaly, need to immediately check if the wiring terminal is loose". Through the Chinese word segmentation tool such as Jieba, the following words are obtained: "component", "E23", "appear", "voltage", "abnormal", "need", "immediately", "check", "wiring terminal", "whether", "loose". Then, the domain-specific semantic dictionary is applied for entity recognition and classification, wherein "component" and "wiring terminal" are identified as component names, "voltage anomaly" is identified as an alarm type, and "check" is identified as an operation behavior phrase. Further, the above identified semantic units are input into the pre-trained BERT model, and the BERT-Chinese model provided by huggingface is used for vector embedding processing, so that "component", "voltage anomaly" and "check" are respectively converted into corresponding 768-dimensional vector representations, for example, "voltage anomaly" is represented as [0.245, -0.138, 0.377,..., 0.099] after embedding. All vectors are stored according to semantic categories, and an index mapping table is established to ensure that the original text segment and the embedded vector correspond one by one. At the same time, a training sample set is constructed for subsequent model use. In the photovoltaic power station work order scenario, each work order description text can be converted into a semantic embedding vector set composed of component words, alarm type words and operation behavior phrases according to the above process, and finally a semantic embedding result set is formed, for example, the record "WO2024061203" contains embedding vectors: [V_component E23], [V_voltage anomaly], [V_check].

[0055] S202: Obtain the image field in the photovoltaic power station work order information set, identify the structural edge, color block anomaly and damage texture area in the image, and extract the state feature vector of the corresponding area through the residual neural network ResNet50 model to obtain the image state feature extraction result.

[0056] Obtain the image field in the photovoltaic power station work order information set. First, read the image file pointed by the field through the OpenCV library of Python. Taking the work order image "IMG_WO2024061203.jpg" as an example, read the image size to pixels, and normalize the RGB three-channel pixel values to interval, then perform structural edge recognition. The core of the Canny edge detection algorithm is to calculate the pixel gradient amplitude of the gray image, and use a double threshold strategy to judge the edge point: set the low threshold to 50 and the high threshold to 150. If the gradient amplitude of a certain pixel point is Therefore, the point is judged as a strong edge; if the gradient of another pixel point is , because and adjacent to a strong edge pixel, it can also be retained; if , it is discarded;

[0057] After the edge pixels in the image are counted, if the number of edge pixels is 6284 and the total number of pixels is , then the structural edge definition is:

[0058] The proportion of edge pixels is ;

[0059] The results show that the structural boundary recognition of the image is clear;

[0060] Subsequently, image color block anomaly analysis is performed, and the K-Means algorithm is used to cluster the pixel vectors in the image. Assuming that the number of clusters is , 4 groups of cluster centers RGB mean values are obtained. Taking the 2nd group as an example, the center is , the normal component surface color standard value is set to , and the Euclidean distance is calculated as follows:

[0061] ;

[0062] Because , it indicates that the color block belongs to an abnormal area;

[0063] Next, image surface damage texture analysis is performed, and the gray level co-occurrence matrix (GLCM) method is used to calculate the texture energy (Energy) feature, which is defined as:

[0064] ;

[0065] Wherein, is the normalized co-occurrence probability of gray pairs in the image. For example, part of the GLCM matrix is as follows:

[0066] , , , …;

[0067] Therefore, the energy is:

[0068] ;

[0069] Because , it indicates that there is a possibility of texture damage in the area;

[0070] Finally, the image is sent into the ResNet50 model to extract a high-dimensional state feature vector, and a 2048-dimensional feature vector is taken from the output before the last fully connected layer Take the first 5 components as an example:

[0071] , , , , …;

[0072] The vector is the state feature representation of the image "IMG_WO2024061203.jpg", which will be spliced with subsequent text semantic embedding, priority level, etc. to form the operation and maintenance work order feature vector table.

[0073] S203: Based on the semantic embedding result and the image state feature extraction result, according to the priority level field in the photovoltaic power station work order information set, the normalization processing and feature splicing are performed, the vector set of text features, image features and level information is integrated, and the operation and maintenance work order feature vector table is constructed;

[0074] Based on the semantic embedding result and the image state feature extraction result, first, read the text embedding vector and image state vector corresponding to each record in the work order information set. Take "WO2024061203" as an example, the text part has been extracted into three semantic units by BERT, including components, alarm types and operation behaviors, each unit is represented by a fixed dimension vector, and the three vectors are spliced into a unified text feature vector in order. At the same time, the image part is output by the ResNet50 model as a fixed dimension image state vector. Then read the priority level field, combine the maximum level and minimum level in all work orders to linearly scale the current value, map the priority to an interval value, and add it as a separate numerical item at the end of the feature. Finally, the text vector, image vector and priority value are spliced to form a complete feature vector. The feature vector is arranged according to the fixed dimension structure, and each work order record generates a unique feature item and is stored as a structured vector table containing key-value pairs. The key is the work order number, and the value is the spliced vector content. The vector table formed by all work orders is summarized as the operation and maintenance work order feature vector table.

[0075] Please refer to Figure 4 , the specific steps of S3 are as follows:

[0076] S301: Obtain the corresponding text vector and image vector in the operation and maintenance work order feature vector table, calculate the cosine similarity between the vectors based on the consistency of the feature space between the semantic embeddings, identify the degree of image-text matching of the work order content, and generate the image-text similarity calculation result.

[0077] After obtaining the text vector and image vector contained in each record in the operation and maintenance work order feature vector table, the two vectors need to be aligned first to ensure that they are in the same vector space, so as to facilitate the semantic similarity measurement between the vectors. The text vector is derived from the multi semantic unit splicing of the BERT model, and its dimension is usually 2304. The image vector is derived from the full connection layer output of the ResNet50 model, and its dimension is 2048. Since the dimensions of the two are inconsistent, mapping and unification processing must be performed. There are two common methods: one is to project the image vector to the dimension space consistent with the text vector through linear transformation, and the other is to expand the shorter vector to the dimension of the longer vector by zero padding. In this embodiment, the second method is adopted, that is, the image vector is padded with zeros at the end to expand it from 2048 dimensions to 2304 dimensions, so as to directly calculate the cosine similarity with the text vector.

[0078] Let the image vector be Then 256 elements with value 0 are added at the end to expand it to: Finally, the image vector with dimension 2304 is obtained , which is consistent with the original text vector.

[0079] The cosine similarity of the two is then calculated, and the formula is:

[0080] ;

[0081] wherein, : represents the text feature vector of a work order, which is a numerical vector extracted from the work order description text by a semantic model (such as BERT), and generally consists of multiple floating point numbers, for example, 2304 dimensions, each element represents the numerical value of the text semantic in a certain feature direction. : represents the image feature vector of the corresponding work order, which is extracted from the image by an image recognition model (such as ResNet50), and is usually a 2048-dimensional floating point vector. If the dimension is inconsistent with , it is expanded to the same dimension by zero padding at the end, and each element represents the response of the image content in a certain feature dimension. : represents the dot product of vectors and , that is, the sum of the corresponding dimension elements after multiplication of the two vectors. The larger the dot product value is, the more consistent the directions of the two vectors are. : represents the L2 norm of vector , that is, the Euclidean length of the vector, which represents the strength or amplitude of the text vector itself. : represents the L2 norm of vector ​L2 norm of the text vector, which is also the length of the image vector, is calculated in the same way . : Cosine similarity, which represents the cosine value of the angle between two vectors in high-dimensional space, ranges from 0 to 1. The closer to 1, the more consistent the direction of the two vectors, that is, the more matched the semantics of the text and image.

[0082] If the text vector , the image vector is extended to ,

[0083] Dot product is: ;

[0084] Lengths are: , ;

[0085] Cosine similarity calculation is: ;

[0086] This completes the alignment of text and image, dimension mapping and image-text matching degree calculation. For each work order, the same processing flow is performed, and finally a set of image-text similarity scores is formed, providing basic feature input for subsequent classification and scoring operations.

[0087] After obtaining the image-text cosine similarity of each work order, the score results need to be divided into intervals and the matching degree is judged to identify whether the image content is consistent with the semantic description of the text. The evaluation standard of matching degree is based on the rules formed after statistical analysis of a large number of historical data samples. The cosine similarity itself is a real number ranging from 0 to 1, and the closer to 1, the more consistent the direction of the text and image vectors, that is, the closer the semantics of the image and the text.

[0088] The specific determination criteria are as follows:

[0089] When the cosine similarity is ≥0.85, it is considered that the image-text is highly matched, the work order description is accurate, and the image content fully reflects the events or status described in the text;

[0090] When the similarity is between 0.65 and 0.85, it is determined to be a moderate match. Such work orders may have image shooting angle, missing area, description briefness, etc. resulting in image-text deviation;

[0091] When the similarity is <0.65, it is considered that the image-text matching degree is low, there are description and image inconsistency, misphotography, invalid image, etc. risks, which need to be strengthened in the subsequent process. Artificial review or system annotation correction;

[0092] ​For example, if the text-image similarity of a work order is 0.92, it is classified as "high match", and if the score of another work order is 0.58, it is identified as "low match". The system records the results for subsequent classification label training and audit risk rating.

[0093] In this way, by unified dimension vector mapping, standardized calculation process and benchmark interval division, the automatic recognition and degree judgment of the consistency of text and image semantics in photovoltaic power station work orders are realized, and the text-image similarity result table is output, providing reliable semantic support for subsequent audit process.

[0094] S302: input the text-image similarity calculation result and the maintenance work order feature vector table into the support vector machine classification model to perform classification identification of the audit state label, and obtain the audit label identification result;

[0095] After splicing the text-image similarity calculation result and the maintenance work order feature vector table, the system generates a complete work order sample feature set, where each record is composed of a text vector (such as 2304 dimensions), an image vector (such as 2048 dimensions, which has been aligned), a priority normalization value (1 dimension), and a text-image similarity score (1 dimension). Finally, a numerical vector structure such as 4354 dimensions is formed as the model training input. The model uses support vector machine (SVM) for audit state label classification identification. Before training, a batch of historical work order samples with labeled audit labels are collected. These labels may include "pass", "need review", and "reject", which correspond to classification labels 0, 1, and 2 respectively during training.

[0096] In the training phase, first, the input vector is subjected to standardization operation, usually using z-score method with mean 0 and variance 1, to avoid model learning bias caused by too large difference in feature values; then the standardized vector and corresponding labels are input to construct a training set, which is sent to the SVM model for fitting. The model tries to construct an optimal hyperplane in the multi-dimensional feature space to maximize the separation between different classes of samples, while introducing soft interval processing for some samples to tolerate a small number of misclassified samples; the model identifies a part of the samples as "support vectors", which are the key points near the classification boundary that determine the separation surface. After the final training, a new sample judgment model for classification is formed.

[0097] During the classification process, the system receives an unlabeled work order data item and constructs it into a 4354-dimensional feature vector using the same method. This feature vector is then fed into a trained SVM model for classification. The model determines which hyperplane segmentation region the vector falls within in feature space and maps it to the corresponding label category. If the sample point is closest to the "pass" category boundary, the model outputs the "pass" label. If it is located in the fuzzy boundary region and close to the review category support vector, the model returns "review required." If it falls significantly away from the positive class support vector, the model outputs the "reject" label.

[0098] Ultimately, each work order is assigned a review tag field to indicate its automated review and classification result. The system then saves this result, binds it to the original work order number, and constructs a structured output table, providing input for the subsequent risk rating module. The entire classification process supports an incremental update mechanism, allowing the training set to be continuously expanded as the number of manual review feedback samples increases, achieving iterative improvements in model performance.

[0099] S303: Construct a scoring input set based on the audit tag recognition results and the priority field and violation feature content in the operation and maintenance work order feature vector table. Then, perform level mapping on the scoring intervals according to the preset risk classification rules. Combine the audit status and risk level to obtain the PV power station audit result.

[0100] After the review label recognition is completed, the system integrates the recognition results with the priority field, image-text similarity score, and violation feature content in the operation and maintenance work order feature vector table to form a scoring input set for subsequent risk assessment calculations. The input items include three types of key variables: one is the review label result, which uses a mapping numerical method to represent the text classification result, such as "passed" is 0.1, "needs review" is 0.6, and "rejected" is 1.0; the second is the priority level field of the work order, which is processed by minimum-maximum normalization and used as a numerical input to reflect the urgency of the task; the third is the violation feature item, which is composed of violation content identified by the rule base matching method. Each violation is divided into different coefficients according to the risk level, and the coefficients are accumulated to form a violation score; at the same time, the image-text similarity score is also introduced, which is the previously calculated result. , as a reverse risk factor for the semantic consistency between text and image.

[0101] The system uses the following risk scoring function to calculate the comprehensive risk value:

[0102] ;

[0103] in, : Final risk score, ranging from 0 to 1; is the audit label risk coefficient, which is used to represent the risk level corresponding to the audit label result given by the machine learning model for the work order. The larger the value, the higher the audit risk level. For example, “pass” is 0.1, “review” is 0.6, and “reject” is 1.0. is the priority normalization value, which represents the standardized numerical value of the urgency of the work order. It is usually mapped to the 01 interval by the minimum-maximum normalization method. The larger the value, the more urgent the work order. is the normalized violation indicator value, which represents the quantitative expression of the violation behavior detected by the system in terms of standardization. First, the number of violation behaviors is counted and combined with their severity levels (such as high, medium, and low) to assign weights. Then, the total violation value is normalized to make it suitable for the unified scoring model. is the image-text similarity score, which is obtained by calculating the cosine similarity between the text vector and the image vector The higher the value, the more consistent the image content and the text description, part of the image-text matching bias is included in the risk calculation. is the audit risk weight coefficient, which is used to adjust the influence of the audit label on the final risk score. Since the audit result comes from the model output, it can directly reflect the model's overall judgment of the work order, so the weight is usually set to medium-high (e.g., 0.3) to highlight its guiding role. is the priority weight coefficient, which is used to measure the role of urgency in risk scoring. Although priority represents the importance of the work order, it does not necessarily represent content risk, so the weight is set to medium (e.g., 0.2). is the violation indicator weight coefficient, which reflects the core role of the violation item in risk composition. Violation content is triggered by rules, directly related to specification, compliance, or technical risk, so its weight should be set to higher (e.g., 0.3) to emphasize its leading role in identifying high-risk work orders. is the image-text bias weight coefficient, which adjusts the influence of image-text content matching bias on the total risk. Since its role is auxiliary judgment, it is usually set to medium-low (e.g., 0.2) to further distinguish risks when other elements are close.

[0104] The setting of the weighting coefficient is mainly based on the intuitive strength relationship of the contribution of each feature item to the overall risk assessment, and follows the principle of "the greater the impact on the risk result, the higher the weight allocation". First, the audit label item represents the direct judgment of the model on the status of the work order processing, which is the output result based on graphics, rules and historical learning, and has strong result orientation, so its weight is set to a high value; second, the violation feature item is derived from rule triggering, which can clearly indicate the potential non-compliance behavior in the work order and has a direct impact on safety and quality, so it is also given a high weight; third, the priority field reflects the urgency or scheduling priority of the work order, although it can explain the importance of timeliness, but its direct effect on content correctness and compliance risk is relatively limited, so the weight is set to medium-low; finally, the graphics-text similarity score is used to judge the consistency of the image and the text description, which belongs to a relative index at the semantic level, although it can reveal the description deviation but will not directly constitute a risk, so the weight of this item is set to an auxiliary medium level. In the overall setting, the sum of the weights of each feature item is 1, ensuring that the score calculation has balance and scalability.

[0105] A work order number "WO2024061203" has an audit label of "Need Review", and its corresponding audit label risk coefficient is ; the priority level is 2, the original level range is 1 to 3, and after normalization, it is ; the system detects two items of violation content, one of which is high risk (weight 1.0) and one of which is medium risk (weight 0.5), with a total violation value of 1.5, and after normalization, it is recorded as ; the cosine similarity between the image and the text of the work order is , and the risk score formula is:

[0106] ;

[0107] Since the final risk score falls into the high-risk interval (defined as ), combined with the audit label "Need Review", the final work order audit result is: Need Review | High Risk.

[0108] Please refer to Figure 5 , the specific steps of S4 are as follows:

[0109] S401: Obtain the audit status and risk level identification in the photovoltaic power station audit result, identify the corresponding text label and level information, parse the audit status text and extract the work order classification keywords, and generate an audit status parsing result;

[0110] After obtaining the audit status and risk level identifier in the photovoltaic power station audit result, the system first performs semantic analysis processing on the audit status text field. Taking the audit result of work order "WO2024061203" as an example, the status label is "need to review | high risk", and the analysis module will decompose it into two parts, identifying "audit status" as "need to review" and "risk level" as "high risk". Then, according to the defined audit label dictionary and risk level mapping table, the system matches the text identifier with the standard label system and corresponds to the internal status code field. At the same time, the semantic content associated with the "need to review" label is extracted, and the decision keywords related to the audit status are identified, such as "boundary sample", "ambiguous judgment", "need to review manually", etc. Then, combined with the classification history and context semantics of the work order itself, the work order attribution category is extracted, such as "inverter problem", "image inconsistency" or "text description ambiguity", etc., to generate a corresponding classification keyword set. The analysis result will be output as an audit status analysis table, each record containing the original text, identified label, corresponding risk level and generated classification keyword set.

[0111] S402: Based on the audit status analysis result and the operation and maintenance work order feature vector table, the keywords of the specified ranked field items are extracted, and the field description text is constructed by combining the fault description field and the annotation content in the image recognition area to obtain the field description generation result;

[0112] After obtaining the audit status analysis result, the system further links the operation and maintenance work order feature vector table, ranks the fields according to the set sorting logic, for example, the top three fields ranked by field influence weight value or information contribution degree are extracted, and the keywords of these fields are extracted. Taking work order "WO2024061203" as an example, the system identifies the top ranked fields as "fault description", "image similarity score" and "violation feature item", which correspond to the text segments and statistical values in the original input. Then, the system performs word segmentation and keyword labeling on the "fault description" field such as "component E23 voltage anomaly", and extracts "component E23" and "voltage anomaly" as the main keywords; then, from the visual annotation information generated by the image recognition area, such as the labels of the detection box coverage area "shadow obstruction" and "loose connection", they are included in the candidate keyword set; finally, the system combines the number of violation feature items and the severity level to obtain the violation level keywords such as "medium risk", "connection error" or "annotation missing", and automatically synthesizes the structural field description text based on the above content, forming an information segment including technical points, identified abnormalities and impact description. The example text is: "Component E23 detects voltage anomaly, image has obstruction and loose connection problems, and is determined as a medium risk work order".

[0113] S403: Based on the field description generation result and the audit state analysis result, arrange the audit state text, risk level description and field content in sequence to generate a structured audit conclusion record;

[0114] After obtaining the field description generation result and the audit state analysis result, the system enters the structured audit conclusion generation process. The results field is spliced and arranged one by one according to the fixed format of "audit state text -> risk level description -> field content information", to generate a structured audit conclusion record that can be directly displayed or archived. For example, combined with the various intermediate data of the work order "WO2024061203", the system first reads the audit state as "need to review", which is used as the first paragraph of the conclusion; then reads the risk level as "high risk", which is mapped to the description text "there are serious inconsistencies between images and compliance abnormalities" according to the preset standard; then the field description result is inserted as the end of the conclusion, including "component E23 detects voltage anomaly, image shows occlusion and wiring problem, violation level is medium", and finally it is integrated into a complete conclusion record: "audit state: need to review; risk level: high risk (there are serious inconsistencies between images and compliance abnormalities); identified problems: component E23 detects voltage anomaly, image exists occlusion and wiring looseness, violation is medium." All work orders generate this type of structured conclusion table according to the same template.

[0115] Please refer to Figure 6 It also includes S5: based on the structured audit conclusion record, optimizing the label consistency and field expression of the audit conclusion content of the operation and maintenance work order, to obtain the optimized audit conclusion content;

[0116] The optimized audit conclusion content includes the label style uniformity item, the semantic expression compression item, and the field order adjustment item.

[0117] The specific steps of S5 are as follows:

[0118] S501: Obtain the audit state text, risk level description and field content in the structured audit conclusion record, identify repeated semantic fragments and redundant field descriptions, and generate an expression compression result;

[0119] After obtaining the audit state text, risk level description, and field content in the structured audit conclusion record, the system first performs semantic compression on the entire text record to identify redundant semantic structures and redundant description fields. For example, the audit conclusion "Audit state: needs review; risk level: high risk (serious inconsistency between image and text and compliance anomaly); identified problem: component E23 detects voltage anomaly, image has occlusion and loose wiring, medium level violation" is processed by the system through text deduplication and semantic vector comparison algorithm. The system identifies that "inconsistency between image and text" and "image has occlusion" have semantic redundancy, and "medium level violation" and "compliance anomaly" are redundant. Then the system uses compression strategy to merge similar phrases, and merges "occlusion and loose wiring" into "image recognition anomaly", and deletes the redundant modification in "compliance anomaly" and shortens it to "medium level violation". After processing, the text of the conclusion is compressed to: "Audit state: needs review", "risk level: high risk (inconsistency between image and text and recognition anomaly)", and "field description: component E23 voltage anomaly, image anomaly, medium level violation". The system compares the difference in the number of words before and after processing, and the length before compression is 56 words, and the length after compression is 37 words, with a compression rate of about 33.9%. The system forms a table record of the original content, compressed text, and compression amplitude information.

[0120] S502: Based on the expression compression result, perform style consistency processing on the key words and label content in the reserved field, map and adjust the description style and semantic boundary according to the label standard, standardize the word range and expression form, and obtain the label unified processing result;

[0121] After obtaining the expression compression result, the system performs style consistency processing on the key words and label content in the reserved field after compression. First, the system identifies the label terms and key words in the text, such as "voltage anomaly", "image anomaly", and "medium level violation", and compares them with the pre-defined label standard items in the label specification table. It is found that "image anomaly" has ambiguous semantics and should be uniformly marked as "image recognition anomaly" according to the standard mapping rules. At the same time, the system standardizes the word order and adjusts the naming style of the key words in different fields, such as unifying "component E23" to "E23 component" and standardizing "medium level violation" to "medium level violation". Then, the system adjusts the words of similar labels appearing in different positions, such as unifying "inconsistency between image and text" to "inconsistency between image and text". The style specification table before and after processing shows that there are 3 inconsistent naming fields in the original text, which are all unified to the standard format after processing, and are updated to the label unified processing result set.

[0122] S503: Based on the label unified processing result, reorganize the field position according to the display order of audit state, risk level, and field description to obtain the optimized audit conclusion content;

[0123] After the label unified processing is completed, the system reorders each field according to the conclusion display structure, and ensures that the finally generated audit conclusion has logical clarity and expression standardization. Taking the processing result "audit state: need to review; risk level: high risk (image and text are inconsistent, image recognition is abnormal); field description: E23 component voltage is abnormal, medium level violation" as an example, the system reorganizes the field position according to the order of audit state first, risk level second, and field description last, and compresses the clause expression to one line of information. The system compares the text structure before and after rearrangement, judges whether the arrangement logic conforms to the preset template, and performs grammar adjustment if the subject is missing or the modifier position is improper. The finally output audit text is "need to review | high risk (image and text are inconsistent, image recognition is abnormal) | E23 component voltage is abnormal, medium level violation", and the output character number is counted and optimized to be controlled within 50 characters to adapt to the interface display standard. All reorganization results are stored in the conclusion display template record, ensuring that the formats of all work order audit results are uniform, prominent and logically complete.

[0124] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any skilled person in the art can use the disclosed technical content to make changes or modifications to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiments without departing from the technical solution content of the present application still belongs to the protection scope of the technical solution of the present application.

Claims

1. A digital system work order review and method based on artificial intelligence, characterized in that: The following steps are involved: S1: Obtain the fault description field and image field of the photovoltaic power station operation and maintenance work order, perform standardization processing, and construct the photovoltaic power station work order information set; The specific steps of S1 are as follows: S101: Obtain the fault description field, equipment number field, priority level field, and on-site image field of the work order recorded on the photovoltaic power station digital platform, uniformly process the format of each type of field, and generate a preliminary field cleaning result; S102: Based on the equipment number field in the preliminary field cleaning result, format matching and naming adjustment are performed according to component standard coding rules to construct a photovoltaic power station work order information set; S2: extracting semantic features of the fault description field and state features of the image field in the photovoltaic power station work order information set, and constructing an operation and maintenance work order feature vector table; The specific steps of S2 are as follows: S201: Obtaining a fault description field in the photovoltaic power station work order information set, identifying component names, alarm types, and operation behavior phrases in the text through semantic parsing, and embedding the identified phrases into trainable vector representations using a bidirectional encoder representation model BERT to generate a semantic embedding result; S202: Obtain an image field from the photovoltaic power station work order information set, identify structural edges, color block anomalies, and damaged texture areas in the image, and extract state feature vectors of the corresponding areas using a residual neural network ResNet50 model to obtain image state feature extraction results; S203: Based on the semantic embedding result and the image state feature extraction result, normalization processing and feature splicing are performed according to the priority level field in the photovoltaic power station work order information set, and a vector set of text features, image features and level information is integrated to construct an operation and maintenance work order feature vector table; S3: Inputting the operation and maintenance work order feature vector table into a support vector machine classification model to identify the audit status of the operation and maintenance work order and classify the corresponding risk level to obtain the photovoltaic power station audit result; The specific steps of S3 are as follows: S301: Obtain the corresponding text vector and image vector in the operation and maintenance work order feature vector table, calculate the cosine similarity between the vectors based on the feature space consistency between the semantic embeddings, identify the degree of image and text matching of the work order content, and generate an image-text similarity calculation result; S302: Input the image-text similarity calculation result and the operation and maintenance work order feature vector table into a support vector machine classification model to perform classification and recognition of the review status label to obtain the review label recognition result; S303: Constructing a scoring input set based on the audit tag recognition result and the priority field and violation feature content in the operation and maintenance work order feature vector table, mapping the scoring intervals to levels according to preset risk classification rules, and combining the audit status and risk level to obtain the PV power station audit result; S4: Integrate the audit status, risk level and corresponding field descriptions in the audit results of the photovoltaic power station to generate a structured audit conclusion record; The specific steps of S4 are as follows: S401: Obtain the audit status and risk level identifier in the audit result of the photovoltaic power station, identify the corresponding text tags and level information, parse the audit status text and extract work order classification keywords, and generate an audit status parsing result; S402: Based on the audit status analysis result and the operation and maintenance work order feature vector table, perform keyword extraction on the field items of the specified ranking, and construct field description text by combining the fault description field and the annotation content in the image recognition area to obtain the field description generation result; S403: Based on the field description generation result and the audit status analysis result, the audit status text, risk level description and field content are arranged in sequence to generate a structured audit conclusion record.

2. The artificial intelligence-based digital system work order review and method according to claim 1 is characterized by: The photovoltaic power station work order information set includes a standardized fault description field, a standardized image field, and an equipment code field. The operation and maintenance work order feature vector table includes a semantic embedding vector, an image status feature vector, and a normalized priority field. The photovoltaic power station audit result specifically includes an audit label value, a risk level identifier, and a graphic consistency score. The structured audit conclusion record includes an audit status text, a risk level description, and a field description.

3. The artificial intelligence-based digital system work order review and method according to claim 1 is characterized by: For the risk level, the formula is used: ； Calculates a risk score between 0 and 1 , divided into corresponding risk levels according to risk scores; in, is the audit label risk factor, is the priority normalized value, is the normalized violation index value, is the audit risk weight coefficient, is the priority weight coefficient, is the violation indicator weight coefficient, is the image-text deviation weight coefficient.

4. The artificial intelligence-based digital system work order review and method according to claim 1 is characterized by: The method further includes S5: optimizing the label consistency and field expression of the audit conclusion content of the operation and maintenance work order based on the structured audit conclusion record to obtain optimized audit conclusion content; The optimized review conclusion content specifically refers to the label style unification item, semantic expression compression item, and field order adjustment item.

5. The artificial intelligence-based digital system work order review and method according to claim 4 is characterized by: The specific steps of S5 are as follows: S501: Obtain the audit status text, risk level description and field content in the structured audit conclusion record, identify repeated semantic segments and redundant field descriptions, and generate an expression compression result; S502: Based on the expression compression result, perform style consistency processing on the keywords and tag content in the retained field, map and adjust the description style and semantic boundary according to the tag standard, standardize the word range and expression form, and obtain a tag unified processing result; S503: Based on the unified processing result of the labels, the field positions are reorganized according to the display order of the audit status, risk level and field description to obtain the optimized audit conclusion content.

Images