Work order cost prediction method and system based on semantic analysis and multi-source data fusion

Abstract

The application provides a work order cost prediction method based on semantic analysis and multi-source data fusion, and belongs to the technical field of artificial intelligence. The method comprises the following steps: obtaining fault text description, equipment parameters, historical maintenance records, spare parts inventory information and environmental data; extracting fault semantic feature vectors through a BERT-base model and an attention mechanism; fusing semantic features and standardized structured data to construct a multi-source feature matrix; predicting work order costs by using a neural network cascade model; and generating maintenance resource allocation strategies according to the predicted values. The application solves the technical problems of low utilization rate of text information and insufficient collaboration of multi-source data in traditional work order cost prediction, and significantly improves the prediction accuracy and maintenance resource allocation efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more particularly to a method and system for predicting work order costs based on semantic analysis and multi-source data fusion. Background Technology

[0002] In industrial production and equipment maintenance, accurate forecasting of work order costs is crucial for cost control and resource planning. Traditional work order cost forecasting methods often have several limitations:

[0003] On the one hand, simply relying on historical cost data from work orders for basic statistical analysis fails to fully explore the deeper semantic information behind the fault descriptions, making it difficult to accurately grasp the complex relationships and inherent patterns inherent in different fault types. For example, when faced with complex fault scenarios, relying solely on surface-level textual descriptions makes it difficult to determine the root cause of the fault and the corresponding handling costs, ignoring key semantic information in the fault descriptions (e.g., "motor coil burnout" implies a higher repair cost than "motor failure"), leading to prediction bias.

[0004] On the other hand, there is a lack of effective integration of diverse data closely related to work order costs, such as equipment operation data and on-site environmental data. Real-time equipment operating status and on-site environmental factors can significantly impact work order processing costs. Traditional methods neglect the integration and utilization of these multi-source data, resulting in prediction models lacking key information inputs and failing to fully reflect the composition logic of actual work order costs, thus affecting prediction accuracy and reliability. Moreover, prediction models often employ single algorithms, which are insufficient for modeling complex nonlinear relationships, especially when the fault description semantics are complex, leading to a significant increase in prediction errors. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method and system for predicting work order costs based on fault semantic analysis and multi-source data fusion. The method involves determining the fault type and its corresponding semantic feature vector through semantic analysis; constructing a multi-source data fusion feature vector by fusing multi-source data; and combining the two to form a work order feature input vector, which is then input into a pre-trained cost prediction model to obtain the cost prediction result. By fusing fault semantic information and multi-source data, this invention can more accurately predict work order costs, providing strong support for enterprise operation and maintenance cost management.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect of the present invention, a method for predicting work order costs based on semantic analysis and multi-source data fusion is provided, characterized by comprising the following steps:

[0008] Acquire work order data, which includes fault description text information and multi-source data associated with the work order. The multi-source data includes at least equipment operation data, historical work order record data, and on-site environmental data.

[0009] The fault description text information is subjected to semantic analysis processing. Natural language processing technology is used to segment the fault description text, perform part-of-speech tagging and named entity recognition, and extract fault keywords and key phrases. According to the preset fault semantic knowledge base, the extracted fault keywords and key phrases are semantically matched and classified to determine the fault type and its corresponding fault semantic feature vector.

[0010] Feature extraction is performed on equipment operation data, historical work order record data and on-site environmental data. Then, the extracted equipment operation status feature parameters, historical work order feature data and on-site environmental feature parameters are fused to construct a multi-source data fusion feature vector.

[0011] By fusing the fault semantic feature vector with the multi-source data fusion feature vector, a multi-source fusion feature matrix is ​​constructed to obtain a complete work order feature input vector;

[0012] The work order feature input vector is input into a pre-trained cost prediction model to obtain the cost prediction result of the current work order, and a corresponding maintenance resource configuration suggestion is generated based on the cost prediction result.

[0013] Preferably, the equipment operation data includes equipment operation parameters, the historical work order record data includes fault text descriptions, historical maintenance records and spare parts inventory information, and the site environment data includes site temperature, humidity, light intensity and site accessibility.

[0014] Preferably, the semantic analysis process includes concatenating all text content in the current work order window to form a text set, processing it through the BERT-base model, extracting the CLS vector of each text as a semantic vector, taking the dimensional average of all text vectors to obtain a vector with a dimension of 768, and generating a unified text feature for the entire window.

[0015] Preferably, the equipment operating status characteristic parameters include voltage, current, and equipment operating time; the historical work order characteristic data includes historical fault handling time, spare parts usage, and manpower input; and the on-site environmental characteristic parameters include on-site temperature, humidity, light intensity, and on-site accessibility.

[0016] Preferably, the multi-source fusion feature matrix is ​​obtained by concatenating the fault semantic feature vector with the standardized structured data into a multi-dimensional feature vector, and then performing feature dimensionality reduction and nonlinear mapping through a fully connected layer.

[0017] Preferably, the cost prediction model is a cascaded neural network model, which determines whether the equipment needs maintenance by constructing an anomaly scoring function. The anomaly scoring function is defined as follows:

[0018] ,

[0019] in, This indicates the anomaly score for device i. The variable representing device i has a risk amplification factor, used to amplify the impact of the mutation process. T represents the operating period of device i. Textual semantic features representing device i's fault The second-order time difference of the state variable, i.e., the acceleration of the change, Indicates time,

[0020] The step of determining whether the current state is abnormal includes:

[0021] When the abnormal state score is greater than a preset threshold, the current state will be marked as abnormal;

[0022] When the abnormal state score is less than a preset threshold, the current state will be marked as normal.

[0023] Preferably, the maintenance resource allocation recommendations include prioritizing work orders based on predicted costs, dynamically recommending the optimal spare parts allocation plan based on spare parts inventory, and generating a visualized maintenance cost analysis report.

[0024] In a second aspect of the invention, a work order cost prediction system based on semantic analysis and multi-source data fusion is provided, characterized in that it comprises, in sequence:

[0025] Data acquisition module: used to collect work order data from multiple sources;

[0026] Semantic analysis module: used to parse fault text descriptions and generate fault semantic feature vectors;

[0027] Data fusion module: used to integrate semantic features and structured data, and output a fused feature matrix;

[0028] Cost prediction module: contains a pre-trained cascaded neural network model that outputs a cost prediction value based on the fused feature matrix;

[0029] Decision support module: Generates maintenance resource allocation strategies based on prediction results.

[0030] The beneficial effects of this invention are as follows:

[0031] By deeply analyzing the semantic information of fault description text, the fault type and its associated characteristics are accurately determined. At the same time, multi-source data such as equipment operating status, historical work order characteristics, and on-site environment are integrated to provide richer and more comprehensive input for the prediction model. This enables the model to more accurately capture the various factors affecting work order costs and their complex relationships. Compared with the traditional simple prediction method based solely on historical costs, this significantly improves the accuracy and reliability of work order cost prediction results, helping enterprises to more accurately plan operation and maintenance budgets and resource allocation.

[0032] It effectively integrates work order-related text data and structured multi-source data, breaking the dependence of traditional methods on a single data source and giving full play to the value of various types of data. This enables enterprises to have more sufficient information basis in data-driven operation and maintenance management decisions. Furthermore, by setting an anomaly scoring function to determine whether the current equipment needs maintenance, it further optimizes the operation and maintenance cost control process and improves the overall operational efficiency of enterprises.

[0033] Accurately predicting work order costs in advance can help companies prepare maintenance funds, rationally allocate technical personnel, and arrange the procurement of maintenance spare parts. This avoids resource waste or maintenance delays caused by inaccurate cost estimates, thereby ensuring the normal operation of equipment, reducing the company's maintenance costs, and enhancing the company's competitive advantage and sustainable development capabilities in the market. Attached Figure Description

[0034] Figure 1 This is a flowchart of the work order cost prediction method based on semantic analysis and multi-source data fusion of the present invention.

[0035] Figure 2 This is a framework diagram of the work order cost prediction system based on semantic analysis and multi-source data fusion of the present invention. Detailed Implementation

[0036] In this embodiment, the hardware used is a high-performance cloud server with the following configuration: an Intel Xeon Gold 6248 processor with a clock speed of 2.5GHz, 128GB of memory, and a 1TB SSD. The operating system is CentOS 7.6. The software programming language is Python 3.7, the database management system is MySQL 8.0, the data analysis tools include Pandas 1.2.0 and NumPy 1.19.2, and the machine learning library is Scikit-learn 0.24.2.

[0037] Please see Figure 1 As shown, in a first aspect of the present invention, a method for predicting work order costs based on semantic analysis and multi-source data fusion is provided, characterized by comprising the following steps:

[0038] Acquire work order data, which includes fault description text information and multi-source data associated with the work order. The multi-source data includes at least equipment operation data, historical work order record data, and on-site environmental data.

[0039] Work order data is extracted from the equipment maintenance management system within the factory and from paper-based repair work orders, including fault description text information and multi-source data associated with the work orders.

[0040] Equipment operation data is collected in real time through the factory's equipment monitoring system. This system is installed on various production equipment and can acquire real-time operating parameters (such as motor speed, temperature, pressure, etc.) and fault alarm information. On-site environmental data is obtained through environmental sensors installed on-site (such as temperature and humidity sensors, light sensors, etc.) and manual on-site inspection records. Historical work order records are stored in the company's internal database, containing detailed information on all equipment maintenance work orders from the past year, as detailed below:

[0041] Fault text description: Obtained in real time from the work order system, example: "On July 3, the frequency converter in Area A frequently tripped due to overload, and the motor made abnormal noises."

[0042] Equipment parameters: extracted from the equipment management database, including rated power (kW), service life (years), and maintenance level (level 1-5);

[0043] Historical maintenance records: Link to the maintenance knowledge base to extract the average maintenance time (hours) and replacement parts cost (yuan) for similar equipment.

[0044] Environmental data: collected via IoT sensors, including temperature (°C), humidity (%), and voltage fluctuation range (±V);

[0045] Data cleaning: Remove duplicate work orders and fill in missing values ​​(environmental data is filled with the mean of a sliding window).

[0046] The fault description text information is subjected to semantic analysis processing. Natural language processing technology is used to segment the fault description text, perform part-of-speech tagging and named entity recognition, and extract fault keywords and key phrases. According to the preset fault semantic knowledge base, the extracted fault keywords and key phrases are semantically matched and classified to determine the fault type and its corresponding fault semantic feature vector.

[0047] The formal processing operates on a minute-level sliding window basis, generating a time slice 't' every minute. Within this window, data from the past 60 seconds is aggregated to form a sample unit. For each time slice, the system performs the following processing steps:

[0048] Error text semantic feature extraction: All text content within the current work order window (including the original post, comments, forwarding instructions, etc.) is concatenated to form a text set. This set is then processed using the BERT-base model to extract the [CLS] vector for each text as a semantic vector. This process is used to generate unified text features for the entire window. The mean pooling of all text vectors yields a vector of dimension 768. This operation is implemented using the transformers library in PyTorch, and it processes an average of 300 to 1000 texts per minute.

[0049] Word segmentation: The Jieba word segmentation tool is used to segment the fault description text. For example, for the fault description text "The motor experiences abnormal high temperature, accompanied by abnormal noise, which may cause equipment shutdown", the word segmentation result is "motor / appears / abnormal / high temperature / , / accompanied / abnormal noise / , / may / cause / equipment / shutdown". Then, part-of-speech tagging is performed on the segmented results to identify nouns (such as "motor", "high temperature", "abnormal noise", "equipment", etc.), verbs (such as "appear", "accompanied by", "cause", etc.), and fault-related entities (such as "motor", "high temperature", "abnormal noise", "equipment shutdown", etc.).

[0050] Semantic matching and classification: The pre-defined fault semantic knowledge base includes the following typical fault types and their semantic feature words and phrase combinations:

[0051] Motor faults: Semantic feature words and phrase combinations include "motor", "high temperature", "abnormal noise", "vibration", "shutdown", etc., and the corresponding semantic feature vector is [1, 0, 1, 1, 0] (assuming the vector dimension is 5, corresponding to 5 categories such as motor, circuit, mechanical transmission, hydraulic, and software faults respectively, and a value of 1 indicates that the feature word or phrase is related to motor faults).

[0052] Circuit fault: Semantic feature words and phrase combinations include "circuit", "short circuit", "open circuit", "voltage", "current", etc., and the corresponding semantic feature vector is [0, 1, 0, 0, 0].

[0053] Mechanical transmission failure: Semantic feature words and phrase combinations include "mechanical transmission", "gear", "bearing", "wear", "jamming", etc., and the corresponding semantic feature vector is [0, 0, 1, 0, 0].

[0054] The fault type is determined by calculating the cosine similarity between keywords and key phrases in the fault description text and semantic feature words and phrase combinations in the knowledge base. For example, for the fault description text mentioned above, the keywords and phrases extracted after word segmentation and part-of-speech tagging include "motor", "abnormal high temperature", "abnormal noise", and "equipment shutdown". The similarity between these keywords and phrases and the semantic feature words and phrase combinations of motor faults is calculated. Assuming the similarity is 0.8, which is higher than the set threshold (e.g., 0.6), the fault is determined to be a motor fault, and the corresponding semantic feature vector is [1, 0, 1, 1, 0].

[0055] Feature extraction is performed on equipment operation data, historical work order record data and on-site environmental data. Then, the extracted equipment operation status feature parameters, historical work order feature data and on-site environmental feature parameters are fused to construct a multi-source data fusion feature vector.

[0056] Equipment operation data feature extraction: Collect statistical data on the operation of all equipment within a given time window to form a feature vector u for the equipment operation data. t ,Include:

[0057] Voltage, current, and equipment operating time.

[0058] All behavioral statistics items are uniformly normalized to their maximum values ​​using the following normalization calculation:

[0059] ,

[0060] in, This represents the original value of the statistical indicator of the i-th device in time slice t. For a tiny positive number Used to avoid division by zero. Normalized Make all dimensions in Interval.

[0061] Historical work order feature extraction: The system tracks maintenance work orders for equipment i within the past year and extracts the following data for the current time window: historical fault handling time, spare parts usage, and manpower input.

[0062] On-site environmental features are extracted by collecting data on on-site temperature, humidity, light intensity, and on-site accessibility during time slice t using sensors.

[0063] Data fusion: This involves fusing characteristic parameters such as equipment runtime, failure frequency, performance indicators of key components, failure handling time of historical work orders, spare parts costs, labor input costs, on-site temperature, humidity, and on-site accessibility. A multi-source data fusion feature vector is constructed using a weighted summation method. The weight of each feature parameter is determined based on its impact on work order costs, obtained through a combination of expert scoring and historical data analysis.

[0064] For example, the weights for equipment runtime (0.1), failure frequency (0.1), key component performance indicators (0.1), historical work order failure processing time (0.2), spare parts cost (0.2), manpower input cost (0.2), site temperature (0.05), humidity (0.05), and site accessibility (0.05) are calculated. After weighted summation, the multi-source data fusion feature vector is obtained as [36, 2, 0.05, 80, 500, 800, 30, 60, 0.8] (values ​​after normalization).

[0065] All the above processing results are integrated into an output structure in the form of a time series:

[0066] ,

[0067] in, This represents a feature vector from multi-source data fusion that incorporates temporal characteristics. Indicates the characteristics of equipment operation data. Indicates characteristics of historical work orders. It indicates the characteristics of the on-site environment.

[0068] Each time slice t forms a triple. .

[0069] By fusing the fault semantic feature vector with the multi-source data fusion feature vector, a multi-source fusion feature matrix is ​​constructed to obtain a complete work order feature input vector;

[0070] semantic features of fault text Feature vector fusion with multi-source data The result of fusion:

[0071] ,

[0072] in This represents the work order feature input vector.

[0073] The feature vector concatenation method of this invention employs direct concatenation, directly combining different feature vectors into a single long vector. This method effectively fuses multi-source data into a unified feature vector, thereby improving the accuracy and reliability of the prediction model.

[0074] The semantic feature vector corresponding to the fault type is concatenated with the multi-source data fusion feature vector [36, 2, 0.05, 80, 500, 800, 30, 60, 0.8] to form a complete work order feature input vector [1, 0, 1, 1, 0, 36, 2, 0.05, 80, 500, 800, 30, 60, 0.8].

[0075] The work order feature input vector is input into a pre-trained cost prediction model to obtain the cost prediction result of the current work order, and a corresponding maintenance resource configuration suggestion is generated based on the cost prediction result.

[0076] The cost prediction model was trained using 1000 historical work order data samples. The training data included work orders with different fault types and cost levels: motor fault work orders accounted for 30%, circuit fault work orders for 25%, mechanical transmission fault work orders for 20%, and other fault types for 25%. The number of neurons in the input layer of the neural network model corresponded to the dimension of the work order feature input vector. There were two hidden layers: the first hidden layer had 20 neurons using the ReLU activation function, and the second hidden layer had 10 neurons using the Sigmoid activation function. The output layer had 1 neuron, used to output the predicted work order cost. The mean squared error loss function was used, and the training iterations were 500 with a learning rate of 0.01.

[0077] This invention also constructs an anomaly scoring function to determine whether equipment needs maintenance, considering whether it is a minor fault that does not affect normal operation, thereby reducing the number of maintenance operations and saving maintenance costs. The anomaly scoring function is defined as follows:

[0078] ,

[0079] in, This indicates the anomaly score for device i. The variable representing device i has a risk amplification factor, used to amplify the impact of the mutation process. T represents the operating period of device i. Textual semantic features representing device i's fault The second-order time difference of the state variable, i.e., the acceleration of the change, Indicates time,

[0080] The steps to determine whether the current state is abnormal include:

[0081] When the abnormal state score is greater than a preset threshold, the current state will be marked as abnormal;

[0082] When the abnormal state score is less than a preset threshold, the current state will be marked as normal.

[0083] like When the value exceeds the preset threshold, the current status will be marked as abnormal and a work order needs to be issued immediately to dispatch an engineer for repair.

[0084] like If the value is less than the preset threshold, the current device status is considered to still be within the time limit. The internal system remains in normal operation and requires no downtime for maintenance.

[0085] The above work order feature input vector is input into the trained neural network model, and the model outputs the predicted cost. The predicted cost is set as... The actual cost of the work order was By calculating the error between the two ,

[0086] ,

[0087] in, The error impact factor is represented by the error correction model parameters:

[0088] Based on the validation results, it was found that the model had slightly larger prediction errors for certain specific fault types (such as mechanical transmission faults). By adjusting the model's hyperparameters, such as increasing the number of neurons in the hidden layer and adding historical work order data samples of mechanical transmission faults for training, the optimized model reduced the average relative error to approximately 3% on the new validation dataset, significantly improving prediction performance and accuracy.

[0089] Ultimately, the maintenance resource allocation recommendations include prioritizing work orders based on predicted costs, dynamically recommending the optimal spare parts allocation plan based on spare parts inventory, and generating a visualized maintenance cost analysis report.

[0090] Obtain information on the types, quantities, and prices of various repair spare parts in the company's inventory, such as motor bearing spare parts, inventory quantity and unit price; motor winding spare parts, inventory quantity and unit price. Also, compile a list of tools and equipment available for repair, such as professional testing instruments and repair tool kits, and record their status (e.g., available, faulty and awaiting repair) and location. For example, a motor fault detector, currently in the available status, is located in the factory's repair workshop tool room.

[0091] Suppose that, based on previous work order cost prediction using fault semantic analysis and multi-source data fusion, the predicted cost for a certain motor fault repair work order is 1500 yuan. Analyzing the cost composition, it is estimated that labor costs account for 40% (approximately 600 yuan), spare parts costs account for 50% (approximately 750 yuan), and tool and equipment usage and other miscellaneous costs account for 10% (approximately 150 yuan).

[0092] Based on the estimated labor costs and the skill matching of the maintenance personnel, it is recommended to dispatch a maintenance team consisting of intermediate-level technician Li Si and junior technician Wang Wu. Li Si has extensive experience in motor maintenance and can lead the maintenance work; Wang Wu will serve as an assistant, responsible for support and learning. It is estimated that the two working together can complete the maintenance task within 5 working days, and the estimated labor costs are consistent with the predicted labor costs.

[0093] Meanwhile, considering Zhang San's high skill level and the fact that his current workload is not saturated, he is listed as a backup maintenance personnel. If complex problems are encountered or technical support is needed during the maintenance process, Zhang San can be deployed in a timely manner.

[0094] Based on the estimated cost of spare parts and the type of failure, it is recommended to prepare 2 spare motor bearings (cost 1000 yuan) and 1 spare motor winding (cost 1000 yuan). The quantity and type of these spare parts are determined according to the spare parts required for common motor failures and the matching of the price of spare parts in stock with the estimated cost of spare parts. This approach can meet maintenance needs while avoiding spare parts waste and excessive inventory costs.

[0095] If spare parts are insufficient in stock, place a purchase order with the supplier in advance to ensure that the spare parts arrive before the maintenance begins.

[0096] If, during the maintenance process, it is found that the originally planned maintenance time may exceed expectations due to some unforeseen factors (such as the complexity of the site environment exceeding expectations), other maintenance personnel (such as Zhang San) should be promptly dispatched to provide support, and the allocation of human resource cost budget should be adjusted accordingly.

[0097] If the actual usage of spare parts is less than the recommended quantity, the remaining spare parts should be properly stored, inventory information updated, and the recommended spare parts quantity optimized in subsequent maintenance resource allocation to improve resource utilization efficiency and cost control accuracy. Meanwhile, comparing the simulated pre-maintenance cost estimates with the cost predictions, the various costs are basically consistent, verifying the rationality of the resource allocation recommendations.

[0098] During the maintenance process, the maintenance progress and resource usage are tracked in real time. This includes factors such as maintenance personnel working hours and the actual number of spare parts used. These are compared and analyzed with the configuration recommendations to identify and adjust any discrepancies promptly.

[0099] This embodiment demonstrates how to generate reasonable maintenance resource allocation suggestions based on cost prediction results, and how to continuously improve the resource allocation plan through verification and optimization to ensure efficient maintenance work and effective cost control.

[0100] In summary, this invention provides an innovative, efficient, and accurate method and system for predicting work order costs in the industrial field, with significant practical application value and broad industry promotion prospects.

[0101] Please see Figure 2 As shown, in a second aspect of the present invention, a work order cost prediction system based on semantic analysis and multi-source data fusion is provided, characterized in that it comprises, in sequence, the following:

[0102] Data acquisition module: used to collect work order data from multiple sources;

[0103] Semantic analysis module: used to parse fault text descriptions and generate fault semantic feature vectors;

[0104] Data fusion module: used to integrate semantic features and structured data, and output a fused feature matrix;

[0105] Cost prediction module: contains a pre-trained cascaded neural network model that outputs a cost prediction value based on the fused feature matrix;

[0106] Decision support module: Generates maintenance resource allocation strategies based on prediction results.

[0107] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A work order cost prediction method based on semantic analysis and multi-source data fusion, characterized in that, Includes the following steps: Acquire work order data, which includes fault description text information and multi-source data associated with the work order. The multi-source data includes at least equipment operation data, historical work order record data, and on-site environmental data. The fault description text information is subjected to semantic analysis processing. Natural language processing technology is used to segment the fault description text, perform part-of-speech tagging and named entity recognition, and extract fault keywords and key phrases. Based on the pre-set fault semantic knowledge base, the extracted fault keywords and key phrases are semantically matched and classified to determine the fault type and its corresponding fault semantic feature vector. Feature extraction is performed on equipment operation data, historical work order record data and on-site environmental data. Then, the extracted equipment operation status feature parameters, historical work order feature data and on-site environmental feature parameters are fused to construct a multi-source data fusion feature vector. By fusing the fault semantic feature vector with the multi-source data fusion feature vector, a multi-source fusion feature matrix is ​​constructed to obtain a complete work order feature input vector; The work order feature input vector is input into a pre-trained cost prediction model to obtain the cost prediction result of the current work order, and a corresponding maintenance resource configuration suggestion is generated based on the cost prediction result. The cost prediction model is a cascaded neural network model that determines whether equipment needs repair by constructing an anomaly scoring function. The anomaly scoring function is defined as follows: ， in, This indicates the anomaly score for device i. The variable representing device i has a risk amplification factor, used to amplify the impact of the mutation process. T represents the operating period of device i. Textual semantic features representing device i's fault The second-order time difference of the state variable, i.e., the acceleration of the change, Indicates time; The steps to determine whether the current state is abnormal include: When the anomaly score is greater than a preset threshold, the current state will be marked as an anomaly; When the abnormal score is less than a preset threshold, the current state will be marked as normal.

2. The work order cost prediction method based on semantic analysis and multi-source data fusion according to claim 1, characterized in that, The equipment operation data includes equipment operation parameters; the historical work order record data includes fault text descriptions, historical maintenance records, and spare parts inventory information; and the on-site environmental data includes on-site temperature, humidity, light intensity, and on-site accessibility.

3. The work order cost prediction method based on semantic analysis and multi-source data fusion according to claim 1, characterized in that, The semantic analysis process includes concatenating all text content in the current work order window to form a text set, processing it through the BERT-base model, extracting the CLS vector of each text as a semantic vector, taking the dimensional average of all text vectors to obtain a vector with a dimension of 768, and generating unified text features for the entire window.

4. The work order cost prediction method based on semantic analysis and multi-source data fusion according to claim 1, characterized in that, The equipment operating status characteristic parameters include voltage, current, and equipment operating time; the historical work order characteristic data includes historical fault handling time, spare parts usage, and manpower input; the on-site environmental characteristic parameters include on-site temperature, humidity, light intensity, and on-site accessibility.

5. The work order cost prediction method based on semantic analysis and multi-source data fusion according to claim 1, characterized in that, The multi-source fusion feature matrix is ​​obtained by concatenating fault semantic feature vectors with standardized structured data into a multi-dimensional feature vector, and then performing feature dimensionality reduction and nonlinear mapping through a fully connected layer.

6. The work order cost prediction method based on semantic analysis and multi-source data fusion according to claim 1, characterized in that, The maintenance resource allocation recommendations include prioritizing work orders based on predicted costs, dynamically recommending the optimal spare parts allocation plan based on spare parts inventory, and generating a visualized maintenance cost analysis report.

7. A work order cost prediction system based on semantic analysis and multi-source data fusion, used to implement the work order cost prediction method based on semantic analysis and multi-source data fusion as described in any one of claims 1 to 6, characterized in that, Including those connected sequentially: Data acquisition module: used to collect work order data from multiple sources; Semantic analysis module: used to parse fault text descriptions and generate fault semantic feature vectors; Data fusion module: used to integrate semantic features and structured data, and output a fused feature matrix; Cost prediction module: contains a pre-trained cascaded neural network model that outputs a cost prediction value based on the fused feature matrix; Decision support module: Generates maintenance resource allocation strategies based on prediction results.

Images