Intelligent inspection system, inspection method, medium and computer equipment

Abstract

The invention provides an intelligent inspection system, an inspection method, a medium and computer equipment. The intelligent inspection system comprises a data monitoring module, an inspection setting module, an exception handling module, an intelligent inspection module, an inspection item management module, a record prediction module, a model training module, a record report module and a user management module. By adopting the intelligent inspection system, the data is pushed to the terminal equipment, the information transmission delay can be reduced, the conventional point location adopts a rechecking mode, and the key point location adopts a blind inspection mode, so that different inspection requirements are flexibly met; the system breaks information islands of all departments and realizes data sharing; abnormal data are analyzed through machine learning, automatic control strategy adjustment is supported, the processing precision and timeliness are improved, and human intervention is reduced; the system covers inspection of multiple aspects such as process parameters, equipment states and environment-friendly energy consumption, production quality can be guaranteed in an all-around mode, and the overall production management level is improved.

Description

Technical Field

[0001] The present invention relates to the field of automation control technology, and more specifically, to an intelligent inspection system, inspection method, medium and computer equipment. Background Art

[0002] In the prior art, the traditional inspection process mainly includes: external operators conduct inspections at a certain frequency, use a card reader to punch in at a specific inspection area, and then use a paper notebook to copy the sensor data. After returning to the office, the data is entered into the corresponding data system. After the external operator finds an abnormal situation, the abnormal situation needs to be reported to the system, and it is determined whether on-site disposal is required, or the corresponding internal operator can deal with the problem.

[0003] This inspection method relies on manual inspection, and the data feedback in the system is delayed, so problems cannot be discovered in time. In addition, for the same abnormal situation, different people will rate the abnormal situation differently, and the treatment methods taken on site cannot be guaranteed to be exactly the same. Therefore, due to different people's different understanding of risks, the abnormal situation cannot be solved in a timely and effective manner. Summary of the invention

[0004] In view of the deficiencies in the prior art, the purpose of the present invention is to provide an intelligent inspection method, device, medium and computer equipment to overcome the shortcomings of the prior art that abnormal conditions cannot be solved promptly and effectively due to different measures taken for abnormal conditions.

[0005] The above technical objectives of the present invention are achieved through the following technical solutions: an intelligent inspection system, comprising:

[0006] A data monitoring module, used to monitor DCS data and push the DCS data to an associated terminal device;

[0007] The inspection setting module is used to adjust the inspection mode according to the level of the point, including: using the verification mode for inspection when the point is a regular point; using the blind inspection mode for inspection when the point is a key point;

[0008] The abnormality handling module is used to record abnormal conditions found during inspection, notify internal operators or external operators to handle them, and structure the abnormal conditions and handling operations into data;

[0009] Intelligent inspection module, used to inspect the production area, including process inspection, equipment inspection and comprehensive inspection;

[0010] The inspection item management module is used to summarize the inspection items in the inspection forms of each department, merge the same inspection items, and reduce repeated inspections;

[0011] A record prediction module, used to record equipment inspection and maintenance records, evaluate the operating status of the equipment based on the inspection and maintenance records, and perform preventive maintenance and predictive maintenance on the equipment based on the operating status of the equipment;

[0012] A model training module, used to train an abnormal condition handling model by using the abnormal condition and handling operation after structured processing as training samples;

[0013] Recording and reporting module, used to enter maintenance and repair records and maintenance records into the data system;

[0014] The user management module is used to manage the roles and permissions of system users.

[0015] An intelligent inspection method is applied to the aforementioned intelligent inspection system, and the intelligent inspection method specifically includes the following steps: using the data monitoring module to monitor DCS data and push the DCS data to the associated terminal device; using the inspection setting module to adjust the inspection mode according to the level of the point, specifically including: using the review mode to inspect when the point is a regular point; using the blind inspection mode to inspect when the point is a key point; using the abnormal handling module to record the abnormal conditions found during the inspection, and notify the internal operator or the external operator to handle, and perform data structured processing on the abnormal conditions and handling operations; using the intelligent inspection module to inspect the production area Inspection, specifically including process inspection, equipment inspection and comprehensive inspection; using the inspection item management module to summarize the inspection items in the inspection forms of each department, merging the same inspection items, and reducing repeated inspections; using the record prediction module to record equipment inspection and maintenance records, evaluate the operating status of the equipment based on the inspection and maintenance records, and perform preventive maintenance and predictive maintenance on the equipment based on the operating status of the equipment; using the model training module to use the abnormal conditions and handling operations after structured processing as training samples to train the abnormal condition handling model; using the record reporting module to enter the maintenance inspection and repair records and maintenance records into the data system; using the user management module to manage the roles and permissions of system users.

[0016] In one embodiment, the intelligent inspection method further includes: using the abnormal condition handling model to analyze the detected abnormal condition and generate a corresponding abnormal condition handling operation.

[0017] In one embodiment, the abnormal condition handling model is used to analyze the detected abnormal condition and generate corresponding abnormal condition handling operations, which specifically include: obtaining abnormal state data of the equipment in the abnormal condition, the abnormal state data including sensor data and a natural language description formed by inspection personnel; performing feature extraction on the sensor data to generate a first feature vector; performing feature extraction on the natural language description to generate a second feature vector; fusing the first feature vector and the second feature vector to generate a fused feature vector; calculating the similarity between the fused feature vector and the abnormal condition feature vector stored in the database; and using the handling operation corresponding to the abnormal condition feature vector having the greatest similarity to the fused feature vector as the abnormal condition handling operation.

[0018] In one embodiment, the extracting features from the sensor data to generate a first feature vector specifically includes: extracting features from the sensor data using a long short-term memory network to generate a first feature vector; extracting features from the natural language description to generate a second feature vector specifically includes: extracting features from the natural language description using a BERT model to generate a second feature vector; fusing the first feature vector and the second feature vector to generate a fused feature vector specifically includes: fusing the first feature vector and the second feature vector to generate a fused feature vector using a fully connected layer; and calculating the similarity between the fused feature vector and the abnormal condition feature vector stored in a database specifically includes: calculating the cosine similarity between the fused feature vector and the abnormal condition feature vector stored in the database.

[0019] In one embodiment, the use of a long short-term memory network to extract features from the sensor data and generate a first feature vector specifically includes: using an input layer of a long short-term memory network to receive sensor data, the sensor data including multiple sensor parameter groups, each of which includes multiple categories of sensor parameters; using the input layer to transmit the received process parameter sequence to a convolution layer for convolution to generate a sensor data feature vector; flattening the sensor data feature vector; adjusting the data shape of the flattened sensor data feature vector according to the input rule of the first LSTM layer; inputting the adjusted sensor data feature vector into the first LSTM layer to learn long-term dependencies, and returning the first hidden state sequence of each time step; inputting the first hidden state sequence into a Dropout layer, and randomly discarding 25% of the data in the first hidden state sequence; inputting the first hidden state sequence after discarding the data into a second LSTM layer to extract time-dependent features, and returning sensor data sequence information corresponding to the hidden state of the last time step; and recording the sensor data sequence information as the first feature vector.

[0020] In one embodiment, the method of using the BERT model to extract features from the natural language description and generate a second feature vector specifically includes: using a word segmenter to split the natural language description to generate a word-gram sequence; adding a head tag at the head of the word-gram sequence and an end tag at the end of the word-gram sequence; replacing each word in the word-gram sequence with an index in a vocabulary to generate an index sequence; using the embedding layer of the BERT model to process each index in the index sequence respectively to generate a corresponding word embedding, position embedding and tag embedding, and calculating the sum of the word embedding, position embedding and tag embedding corresponding to each index to generate a total embedding; using the Transform encoder layer of the BERT model to encode each total embedding to generate a word-gram vector corresponding to each total embedding, and a second feature vector associated with the head tag.

[0021] In one embodiment, the step of fusing the first feature vector and the second feature vector using a fully connected layer to generate a fused feature vector specifically includes:

[0022] The step of concatenating the first feature vector and the second feature vector to generate a concatenated vector includes:

[0023] concatenation vector = [first eigenvector; second eigenvector];

[0024] Among them, the dimension of the concatenated vector is n1+n2, n1 is the dimension of the first eigenvector, and n2 is the dimension of the second eigenvector;

[0025] The concatenated vector is input into the fully connected layer for fusion to generate a fused feature vector, specifically including:

[0026] Fusion feature vector = Tanh (concatenated vector W + b);

[0027] Among them, W is the weight matrix with dimension (n1+n2)×m, m is the dimension of the output of the fully connected layer; b is the bias vector with dimension m; Tanh is the activation function.

[0028] A computer-readable storage medium stores a computer program, which implements the steps of the above method when executed by a processor.

[0029] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0030] In summary, the present invention has the following beneficial effects: an intelligent inspection system, comprising: a data monitoring module, an inspection setting module, an abnormality handling module, an intelligent inspection module, an inspection item management module, a record prediction module, a model training module, a record reporting module, and a user management module; the intelligent inspection system of the present invention is used to push real-time data to the mobile devices of on-site inspectors, so that production and equipment status can be obtained instantly, information transmission delays can be reduced, and response speed can be improved. Conventional points adopt a review mode, and key points adopt a blind inspection mode, which can flexibly adapt to different inspection needs; the system breaks down information silos of various departments, realizes data sharing, promotes information exchange between departments, and improves the efficiency of collaborative work; analyzes abnormal data through machine learning, supports automatic adjustment of control strategies, improves the accuracy and timeliness of disposal, and reduces human intervention; the system covers multiple inspections such as process parameters, equipment status, environmental protection and energy consumption, and can comprehensively guarantee production quality and improve the overall production management level. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a structural diagram of an intelligent inspection system of the present invention;

[0032] Figure 2 This is a flow chart of an intelligent inspection method of the present invention;

[0033] Figure 3 This is a flow chart of an intelligent inspection method of the present invention;

[0034] In the figure, 1. Data monitoring module; 2. Inspection setting module; 3. Abnormal handling module; 4. Intelligent inspection module; 5. Inspection item management module; 6. Record prediction module; 7. Model training module; 8. Record report module; 9. User management module. DETAILED DESCRIPTION

[0035] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings. Several embodiments of the present invention are shown in the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein.

[0036] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal connection of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. The terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0037] In the present invention, unless otherwise clearly specified and limited, the first feature being "above" or "below" the second feature may include the first and second features being in direct contact, or may include the first and second features being in contact not directly but through another feature between them. Moreover, the first feature being "above", "above" and "above" the second feature includes the first feature being directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. The first feature being "below", "below" and "below" the second feature includes the first feature being directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature. The terms "vertical", "horizontal", "left", "right", "above", "below" and similar expressions are for illustrative purposes only, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0038] Embodiment 1

[0039] In order to solve the above problems, the present invention provides an intelligent inspection system, such as Figure 1 As shown, including:

[0040] The data monitoring module is used to monitor DCS (distributed control system) data and push the DCS data to the associated terminal device; the data monitoring module pushes the real-time data of the DCS (distributed control system) to the mobile devices of on-site inspectors, which can instantly obtain the production and equipment status, reduce information transmission delays, and improve response speed.

[0041] The inspection setting module is used to adjust the inspection mode according to the level of the point, including: using the review mode for inspection when the point is a regular point; using the blind inspection mode for inspection when the point is a key point; the inspection setting module supports the review mode and the blind inspection mode, and selects the appropriate inspection mode according to the importance of different points. The review mode is used for regular points, and the blind inspection mode is used for key points, which can flexibly adapt to different inspection needs.

[0042] The abnormality handling module is used to record abnormal conditions found during the inspection, notify internal operators or external operators to handle them, and perform data structured processing on abnormal conditions and handling operations; the abnormality handling module performs structured data processing on abnormal conditions found during the inspection process, and performs knowledge management on handling measures, which is helpful for experience accumulation and process optimization;

[0043] The intelligent inspection module is used to inspect the production area, including process inspection, equipment inspection and comprehensive inspection. The intelligent inspection module can combine DCS real-time monitoring and on-site inspection data to quickly identify safety hazards and process anomalies to ensure the safety and stability of the production process. The inspection targets cover process parameters, equipment status, environmental protection and energy consumption, etc., which can comprehensively guarantee production quality and improve the overall production management level.

[0044] The inspection item management module is used to summarize the inspection items in the inspection forms of each department, merge the same inspection items, and reduce duplicate inspections; the inspection item management module integrates and allocates the inspection items of different departments such as safety, production, equipment, fire protection, and environmental protection to avoid duplicate inspections, optimize the inspection process, and greatly improve the work efficiency of on-site inspectors. It can also break the information islands of each department, realize data sharing, promote information exchange between departments, and improve the efficiency of collaborative work.

[0045] The record prediction module is used to record the equipment inspection and maintenance records, evaluate the operating status of the equipment based on the inspection and maintenance records, and perform preventive maintenance and predictive maintenance on the equipment based on the operating status of the equipment; by performing preventive maintenance and predictive maintenance on the equipment, the probability of equipment failure can be reduced, and by evaluating the operating status of the equipment, frequent maintenance that affects the normal operation of the equipment can be avoided, thereby balancing the correlation between equipment failure and efficiency.

[0046] The model training module is used to train the abnormal situation handling model by using the abnormal conditions and handling operations after structured processing as training samples; through data modeling and machine learning, the system can gradually realize automated process control, move towards the unmanned operation goal of the "black screen factory", and promote the intelligent transformation and upgrading of the factory. After the handling measures are structured and trained through machine learning, the control parameters can be automatically adjusted to reduce human intervention and achieve a certain degree of automated operation.

[0047] The record reporting module is used to enter maintenance inspection and repair records and maintenance records into the data system to facilitate future verification.

[0048] The user management module is used to manage the roles and permissions of system users to prevent low-privilege roles from mistakenly operating the system, causing security risks and losses.

[0049] In practical applications, the parameters that need to be detected through the inspection process mainly include two categories: 1. Process; 2. Equipment.

[0050] The variables that need to be monitored in process inspection usually include: temperature, pressure, liquid level, and flow rate. The objects of process inspection mainly include: static equipment, such as chemical reaction towers, tanks, kettles, etc. The types of problems in process inspection mainly include: safety risks, quality problems, and energy consumption problems caused by process fluctuations. When there are risks, the disposal methods mainly include: when the operating data deviates from the upper and lower limits and process fluctuations occur, optimization is carried out by optimizing and adjusting the operating parameters, such as adjusting the valve opening. When the process data reaches the critical point, the system will alarm and take emergency measures if necessary, such as cutting off or closing the valve. The goal of process inspection is to optimize process operation through automatic control, reduce fluctuations, and ensure safety and efficiency.

[0051] The variables that need to be monitored for equipment inspection mainly include: spectrum monitoring of temperature, vibration, displacement, current, and voltage. The objects of equipment inspection mainly include: dynamic equipment, such as compressors, circulating water pumps, motors, etc. Predictive maintenance: When the equipment detects an unhealthy state (such as abnormal temperature or vibration), it may cause a failure. The system recommends emergency parking and repair to avoid more serious problems. This is similar to the "illness" of the equipment. Preventive maintenance and care: After a long period of operation, the equipment may become fatigued and require regular maintenance and care to prevent potential failures. This is equivalent to "nutritional health care" for the equipment to extend the service life of the equipment and ensure continued efficient operation.

[0052] Embodiment 2

[0053] In one embodiment, the present invention provides an intelligent inspection method, which is applied to the aforementioned intelligent inspection system. The intelligent inspection method specifically includes the following steps:

[0054] S1. Use the data monitoring module to monitor DCS data and push the DCS data to the associated terminal device; the DCS data can be pushed to the associated terminal device in real time through the data monitoring module to realize automatic data monitoring, reduce human intervention, improve the transparency and real-time performance of the production process, and facilitate inspection personnel to make quick responses.

[0055] S2. Utilize the inspection setting module to adjust the inspection mode according to the level of the point, specifically including: adopting the verification mode for inspection when the point is a regular point; adopting the blind inspection mode for inspection when the point is a key point; the inspection setting module realizes flexible management and targeted inspection of points of different levels, improves inspection efficiency and accuracy, especially at key points, avoids human interference, and ensures the objectivity of data.

[0056] S3. Use the exception handling module to record abnormal conditions found during inspection, notify internal operators or external operators to handle them, and perform data structuring on the abnormal conditions and handling operations; the exception handling module records and structures abnormal conditions to ensure that abnormal information can be processed and fed back in a timely manner. The data structuring provides a basis for subsequent analysis. At the same time, the system automatically notifies internal and external operators to handle the situation, shortening the response time and improving production safety and problem handling efficiency.

[0057] S4. Use the intelligent inspection module to inspect the production area, including process inspection, equipment inspection and comprehensive inspection. The intelligent inspection module provides process, equipment and comprehensive inspection, covering multi-dimensional inspection items in the production process. Through intelligent means, the inspection scope is wide and efficient, ensuring comprehensive monitoring of process parameters and equipment status in the production process.

[0058] S5. Use the inspection item management module to summarize the inspection items in the inspection forms of each department, merge the same inspection items, and reduce duplicate inspections; the inspection item management module merges similar inspection items to reduce duplicate inspection items between different departments, thereby improving inspection efficiency, optimizing resource allocation, avoiding redundant operations, and improving overall management efficiency.

[0059] S6. Using the record prediction module to record the equipment inspection and maintenance records, evaluating the equipment's operating status based on the inspection and maintenance records, and performing preventive maintenance and predictive maintenance on the equipment based on the equipment's operating status; using the record prediction module to perform preventive maintenance and predictive maintenance on the equipment based on the equipment inspection and maintenance records and operating status evaluation, effectively reducing the equipment's unplanned downtime and maintenance costs; the prediction and evaluation of the equipment's operating status provide data support for equipment maintenance, reduce the blindness and uncontrollability of maintenance, and improve the level of equipment management.

[0060] S7. Use the model training module to take the abnormal conditions and handling operations after structured processing as training samples to train the abnormal condition handling model; the model training module takes the abnormal conditions and handling operation data after structured processing as training samples to form an abnormal condition handling model, so that the system can automatically generate optimized handling plans when similar abnormalities occur in the future, thereby realizing intelligent fault management and decision support.

[0061] S8. Use the record reporting module to enter the maintenance inspection and repair records and maintenance records into the data system; enter the maintenance inspection and repair records and maintenance records into the system to form a systematic data management and knowledge base, which is convenient for the review and analysis of historical data, and helps to optimize maintenance strategies and improve management levels.

[0062] S9. Use the user management module to manage the roles and permissions of system users; by managing the roles and permissions of system users, ensure the security of data and system, and at the same time realize the control and management of different user operation permissions, thereby improving the security and controllability of the system.

[0063] To sum up, the intelligent inspection method of the present application provides an automated, intelligent, flexible and efficient production inspection and management system by organically combining modules such as data monitoring, intelligent inspection, abnormal handling, preventive maintenance and predictive maintenance, and user management. It greatly improves the safety, stability and efficiency of the production process, and provides strong technical support for enterprise intelligent manufacturing and intelligent operation and maintenance.

[0064] In one embodiment, the intelligent inspection method further includes: using the abnormal condition handling model to analyze the detected abnormal condition and generate a corresponding abnormal condition handling operation.

[0065] In actual applications, when an abnormal situation is detected, the system can automatically analyze the abnormal situation and generate corresponding disposal operation suggestions without human intervention. This greatly improves the automation level of exception handling and reduces the subjectivity and delay caused by human judgment. The disposal operations generated by model analysis can be standardized and standardized to ensure the consistency and efficiency of abnormal disposal and reduce the uncertainty in the disposal process. By continuously improving the model, the system can solidify expert experience and best practices in the form of data, and novice operators can also use the system to obtain decision-making support equivalent to that of senior personnel, realizing knowledge inheritance and sharing.

[0066] In one embodiment, the abnormal condition handling model is used to analyze the detected abnormal condition and generate corresponding abnormal condition handling operations, which specifically include: obtaining abnormal state data of the equipment in the abnormal condition, the abnormal state data including sensor data and a natural language description formed by inspection personnel; performing feature extraction on the sensor data to generate a first feature vector; performing feature extraction on the natural language description to generate a second feature vector; fusing the first feature vector and the second feature vector to generate a fused feature vector; calculating the similarity between the fused feature vector and the abnormal condition feature vector stored in the database; and using the handling operation corresponding to the abnormal condition feature vector having the greatest similarity to the fused feature vector as the abnormal condition handling operation.

[0067] Specifically, this embodiment fuses the feature vectors of sensor data and natural language description to generate richer fused feature vectors, which can more comprehensively and accurately reflect the actual abnormal state of the equipment and avoid the analysis bias caused by a single data source; based on similarity matching, it can quickly find the historical case closest to the current abnormality and give the corresponding disposal plan, which reduces the time of manual analysis and judgment and improves the efficiency of abnormality handling; as the system usage time increases, the system will accumulate more and more abnormal cases and disposal methods, continuously enrich and improve the abnormal condition database, so that the system can provide more accurate and effective disposal suggestions when dealing with complex and diverse abnormalities; combining natural language processing with automated decision-making, the system can not only automatically process objective data, but also understand and utilize subjective information input manually, thereby realizing intelligent analysis and decision support in more complex scenarios; through automated analysis and decision-making suggestions, the necessity of manual participation and intervention is greatly reduced, and the automation level of the system is further improved.

[0068] In one embodiment, the extracting features from the sensor data to generate a first feature vector specifically includes: extracting features from the sensor data using a long short-term memory network to generate a first feature vector; extracting features from the natural language description to generate a second feature vector specifically includes: extracting features from the natural language description using a BERT model to generate a second feature vector; fusing the first feature vector and the second feature vector to generate a fused feature vector specifically includes: fusing the first feature vector and the second feature vector to generate a fused feature vector using a fully connected layer; and calculating the similarity between the fused feature vector and the abnormal condition feature vector stored in a database specifically includes: calculating the cosine similarity between the fused feature vector and the abnormal condition feature vector stored in the database.

[0069] Specifically, the long short-term memory network (LSTM) is good at processing time series data and can capture the time dependency and dynamic change characteristics in sensor data. Sensor data is often continuous time series data (such as temperature, pressure, vibration, displacement, etc.). LSTM can effectively extract the implicit patterns and features and generate the first feature vector representing the process and equipment status. The BERT model processes natural language data and can extract deep semantic features from natural language descriptions. The natural language description of the inspector may contain the details and perception information of the equipment status (such as "the equipment has abnormal noise" and "the vibration frequency is abnormal", etc.). The BERT model can be used to convert these descriptions into the second feature vector to capture the semantic information. The first feature vector generated by LSTM and the second feature vector generated by BERT are fused using the fully connected layer, which helps to comprehensively process the time series features and semantic features. The fused feature vector contains the dynamic status information of the equipment and the perception information of the inspector, and has higher expression ability. The cosine similarity calculation is simple and efficient, and can quickly compare the similarity between the fused feature vector and the abnormal condition feature vector stored in the database, facilitating real-time online abnormal condition identification and disposal operation recommendations. In the feature fusion process, even if one type of data (such as sensor data or natural language description) is incomplete or noisy, the other type of data can still provide useful information, thereby improving the system's fault tolerance and overall robustness to missing or abnormal data.

[0070] In one embodiment, the use of a long short-term memory network to extract features from the sensor data and generate a first feature vector specifically includes: using an input layer of a long short-term memory network to receive sensor data, the sensor data including multiple sensor parameter groups, each of which includes multiple categories of sensor parameters; using the input layer to transmit the received process parameter sequence to a convolution layer for convolution to generate a sensor data feature vector; flattening the sensor data feature vector; adjusting the data shape of the flattened sensor data feature vector according to the input rule of the first LSTM layer; inputting the adjusted sensor data feature vector into the first LSTM layer to learn long-term dependencies, and returning the first hidden state sequence of each time step; inputting the first hidden state sequence into a Dropout layer, and randomly discarding 25% of the data in the first hidden state sequence; inputting the first hidden state sequence after discarding the data into a second LSTM layer to extract time-dependent features, and returning sensor data sequence information corresponding to the hidden state of the last time step; and recording the sensor data sequence information as the first feature vector.

[0071] In practical applications, the input layer does not perform any processing after receiving the process parameter sequence, and directly passes the process parameter sequence to the convolution layer for convolution; the convolution layer can extract the short-term local features of the sensor data and enhance the model's perception of sudden anomalies. The convolution operation is very effective in capturing the patterns of sensor data in short-term time changes and can identify short-term trends and fluctuations.

[0072] In the present application, as described in Example 1, the data of the sensor mainly includes two types: process parameters and equipment parameters, among which the process mainly includes: parameters temperature, pressure, liquid level, flow rate. Equipment parameters mainly include: spectrum monitoring of temperature, vibration, displacement, current, and voltage. These parameters can be arbitrarily combined to form a corresponding process parameter sequence, or a parameter sequence including all types can be set. When a parameter of a certain type is not abnormal, the parameter is replaced by 0 to improve the versatility of the parameter sequence.

[0073] The flattening layer is used to flatten the sensor data feature vector to generate the corresponding one-dimensional vector. Converting high-dimensional convolutional features into one-dimensional vectors facilitates processing by the subsequent fully connected layer, but the flattening operation will destroy the structural information of the time series, which may affect the model's capture of time dependencies in the subsequent LSTM layer. Therefore, before inputting into the first LSTM layer, the flattened one-dimensional vector needs to be resized to the dimension required by the first LSTM layer.

[0074] Specifically, the hidden units in the first LSTM layer receive the sensor data feature vectors respectively, and update the state through the forget gate, input gate, output gate and memory unit in the first LSTM layer. In the case of the current time step, the LSTM layer receives the input x of the current time step t , the hidden state h of the previous time step t-1 and the memory cell state C t-1 ; Multiple hidden units process the same input information in parallel.

[0075] Since multiple hidden units are independent of each other, the data processing process mainly includes:

[0076] The forget gate activation value of the jth hidden unit in the tth time step Specifically:

[0077]

[0078] Among them, σ represents the Sigmoid activation function; is the forget gate weight matrix of the jth hidden unit; is the bias term, the activation value of the forget gate Used to control the information that needs to be forgotten in the state of the memory unit.

[0079] The input gate activation value of the jth hidden unit at the tth time step Specifically:

[0080]

[0081] Among them, σ represents the Sigmoid activation function; is the input gate weight matrix of the jth hidden unit; is the bias term, the input gate activation value Used to control which new information in the current time step t needs to be written into the memory cell.

[0082] In the tth time step, the jth hidden unit calculates its candidate memory unit state

[0083]

[0084] Use the candidate memory state to update the memory unit state of the jth hidden unit at the previous time step t-1 Get the memory cell state at the current time step:

[0085]

[0086] The memory cell state means that each hidden unit can independently decide what information to keep or forget in its memory cell state.

[0087] The output gate activation value of the jth hidden unit at the tth time step Specifically:

[0088]

[0089] Among them, σ represents the Sigmoid activation function; is the output gate weight matrix of the jth hidden unit; is the bias term, the output gate activation value Used to control the information output from the memory unit.

[0090] In the tth time step, the jth hidden unit calculates the hidden state of the current time step t, including:

[0091]

[0092] The adjusted process parameter feature vector is outputted by all hidden units at each time step. t, combined into an output vector, which is the first hidden state sequence; in order to avoid overfitting of the model, the first hidden state sequence needs to be input into the Dropout layer to randomly discard 25% of the data to improve the generalization ability of the model. The discarding process keeps the shape of the vector unchanged, but some data becomes 0. Then the first hidden state sequence after the discarded data is input into the second LSTM layer for processing, and finally outputs the sensor data sequence information output at the last time step, in order to summarize the information of the entire time series and provide input for the subsequent fully connected layer.

[0093] In one embodiment, the method of using the BERT model to extract features from the natural language description and generate a second feature vector specifically includes: using a word segmenter to split the natural language description to generate a word-gram sequence; adding a head tag at the head of the word-gram sequence and an end tag at the end of the word-gram sequence; replacing each word in the word-gram sequence with an index in a vocabulary to generate an index sequence; using the embedding layer of the BERT model to process each index in the index sequence respectively to generate a corresponding word embedding, position embedding and tag embedding, and calculating the sum of the word embedding, position embedding and tag embedding corresponding to each index to generate a total embedding; using the Transform encoder layer of the BERT model to encode each total embedding to generate a word-gram vector corresponding to each total embedding, and a second feature vector associated with the head tag.

[0094] In practical applications, before inputting text into the BERT model, the text needs to be properly preprocessed to meet the input requirements of the model. First, the original text string needs to be converted into a token sequence that the model can process. Specifically, in this embodiment, the WordPiece tokenizer is used to split it into smaller sub-word units. The tokenizer can handle unregistered words and rare words. For example, the original text is "The temperature continues to rise during the operation of the equipment, the pressure is also increasing, and there is abnormal noise."

[0095] The word segmentation results are:

[0096] "During the operation of the equipment, the temperature continued to rise, and the pressure was also increasing, and there were abnormal noises."

[0097] Add special tags, including the beginning tag and the end tag. The beginning tag is specifically the [CLS] tag: the beginning tag of the sentence, located at the first position of the sequence, representing the aggregate information of the entire sentence. The end tag is specifically the [SEP] tag: the end tag or separator of the sentence, used to distinguish different sentences. Map each word to a unique index in the vocabulary, and you can get the index sequence:

[0098] “[101,6432,8154,756,6121,704,7444,2421,677,1297,7770,7770,8024,4263,1217,738,1762,6612,1920,8024,3300,7368,2496,1715,4189,511,102]”

[0099] Among them,

[101] is the beginning mark and

[102] is the end mark.

[0100] The index sequence is input into the BERT model, and each word-meta index is mapped to a corresponding word embedding (Token Embeddings) using the input embedding layer of the BERT model. In this embodiment, the dimension of the word embedding is 768 dimensions, and a position embedding (Position Embeddings) is generated according to each word-meta position: it represents the position information of the word-meta in the sequence, and helps the model recognize the word order. According to the number of sentences input, a token type embedding (Token Type Embeddings) is generated for each word-meta to distinguish different sentences. When only one sentence is input, the token type embeddings are all 0. The total embedding is calculated based on the three embeddings, and the total embedding = word embedding + position embedding + token type embedding.

[0101] In this embodiment, the BERT model includes 12 layers of Transformer encoders, each layer includes a multi-head self-attention mechanism and a feedforward neural network; the self-attention mechanism allows the model to encode a certain word while paying attention to other words in the sequence to capture the global context. A residual connection layer is also provided between the Transformer encoders to enable the model to be trained at a deeper level and alleviate the gradient vanishing problem. The results output by the Transformer encoder include: Sequence Output: The shape is [sequence length, hidden dimension], and each word has a corresponding output vector. And the output of the [CLS] tag: located in the first position of the sequence, its output vector is designed to represent the aggregate information of the entire sequence. Among them, the output of the [CLS] tag is the second feature vector in this embodiment; the output vector of the [CLS] tag is used to represent the semantic information of the entire input sequence; the [CLS] vector contains a comprehensive semantic understanding of the entire input text, and the length of the [CLS] vector is fixed. Regardless of the length of the input text, the dimension of the [CLS] vector is fixed, which is convenient for fusion with other features or input into downstream models. Compared with pooling or attention mechanism for the entire sequence, the [CLS] vector provides a simple and efficient method.

[0102] In one embodiment, the step of fusing the first feature vector and the second feature vector using a fully connected layer to generate a fused feature vector specifically includes:

[0103] The step of concatenating the first feature vector and the second feature vector to generate a concatenated vector includes:

[0104] concatenation vector = [first eigenvector; second eigenvector];

[0105] Among them, the dimension of the concatenated vector is n1+n2, n1 is the dimension of the first eigenvector, and n2 is the dimension of the second eigenvector;

[0106] The concatenated vector is input into the fully connected layer for fusion to generate a fused feature vector, specifically including:

[0107] Fusion feature vector = Tanh (concatenated vector W + b);

[0108] Among them, W is the weight matrix with dimension (n1+n2)×m, m is the dimension of the output of the fully connected layer; b is the bias vector with dimension m; Tanh is the activation function.

[0109] In practical applications, the fully connected layer effectively combines features from two different sources to improve the model's ability to understand complex tasks; during the weight training process of the fully connected layer, it can automatically learn which features are more important, thereby increasing the weight of useful information and suppressing useless information; by properly setting the output dimension of the fully connected layer, high-dimensional features can be reduced in dimension to reduce computational complexity. Through the above steps, the first eigenvector and the second eigenvector are effectively fused into a new feature representation, which can enhance the performance and robustness of the model when processing complex tasks. Since Tanh's output is between [-1,1], it can accelerate the convergence process and improve the training efficiency of the model.

[0110] In one embodiment, calculating the cosine similarity between the fused feature vector and the abnormal condition feature vector stored in the database specifically includes:

[0111]

[0112] Among them, A represents the fused feature vector, B represents the abnormal condition feature vector; ||A|| represents the modulus of the fused feature vector; ||B|| represents the modulus of the abnormal condition feature vector.

[0113] in,

[0114]

[0115] In this embodiment, the result range of the cosine similarity is between [-1, 1]. When the cosine similarity of two vectors is closer to 1, the similarity between the two vectors is higher. When the cosine similarity of two vectors is closer to -1, the similarity between the two vectors is lower.

[0116] In summary, the intelligent inspection method of the present application, by using the long short-term memory network to capture the time dependency and dynamic change characteristics in the sensor data, generates a first feature vector representing the device status. By using the BERT model to process natural language data, it is possible to extract deep semantic features from the natural language description and convert them into a second feature vector. The first feature vector generated by LSTM and the second feature vector generated by BERT are fused using a fully connected layer, which helps to comprehensively process the time series features and semantic features. The fused feature vector contains the dynamic state information of the equipment and the perception information of the inspectors, and has higher expressive power. In the process of feature fusion, even if one type of data is incomplete or noisy, the other type of data can still provide useful information, thereby improving the system's fault tolerance and overall robustness to missing or abnormal data.

[0117] Embodiment 3

[0118] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the intelligent inspection method as described in the second embodiment is implemented.

[0119] Embodiment 4

[0120] A computer device, which may be a server, may have an internal structure as shown in Figure 3 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. When the computer program is executed by the processor, an intelligent inspection method is implemented.

[0121] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0122] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0123] Using the data monitoring module to monitor DCS data, and pushing the DCS data to an associated terminal device;

[0124] The inspection setting module is used to adjust the inspection mode according to the level of the point, specifically including: using the verification mode for inspection when the point is a regular point; using the blind inspection mode for inspection when the point is a key point;

[0125] The abnormality handling module is used to record abnormal conditions found during inspection, notify internal operators or external operators to handle them, and perform data structured processing on the abnormal conditions and handling operations;

[0126] The intelligent inspection module is used to inspect the production area, including process inspection, equipment inspection and comprehensive inspection;

[0127] The inspection item management module is used to summarize the inspection items in the inspection forms of each department, merge the same inspection items therein, and reduce repeated inspections;

[0128] Using the record prediction module to record equipment inspection and maintenance records, evaluating the operating status of the equipment based on the inspection and maintenance records, and performing preventive maintenance and predictive maintenance on the equipment based on the operating status of the equipment;

[0129] Using the model training module to take the abnormal conditions and handling operations after structured processing as training samples, training the abnormal condition handling model;

[0130] Using the record reporting module to enter the maintenance and repair records and maintenance records into the data system;

[0131] The user management module is used to manage the roles and permissions of system users.

[0132] In one embodiment, the intelligent inspection method further includes: using the abnormal condition handling model to analyze the detected abnormal condition and generate a corresponding abnormal condition handling operation.

[0133] In one embodiment, the abnormal situation handling model is used to analyze the detected abnormal situation and generate a corresponding abnormal situation handling operation, specifically including:

[0134] Acquire abnormal state data of the device in an abnormal condition, wherein the abnormal state data includes sensor data and a natural language description formed by an inspection personnel;

[0135] Performing feature extraction on the sensor data to generate a first feature vector;

[0136] Performing feature extraction on the natural language description to generate a second feature vector;

[0137] Fusing the first feature vector and the second feature vector to generate a fused feature vector;

[0138] Calculating the similarity between the fused feature vector and the abnormal condition feature vector stored in the database;

[0139] The handling operation corresponding to the abnormal situation feature vector having the greatest similarity to the fused feature vector is used as the abnormal situation handling operation.

[0140] In one embodiment, extracting features from the sensor data to generate a first feature vector specifically includes: extracting features from the sensor data using a long short-term memory network to generate a first feature vector;

[0141] The extracting features from the natural language description to generate a second feature vector specifically includes: extracting features from the natural language description using a BERT model to generate a second feature vector;

[0142] The fusing the first feature vector and the second feature vector to generate a fused feature vector specifically includes: using a fully connected layer to fuse the first feature vector and the second feature vector to generate a fused feature vector;

[0143] The calculating the similarity between the fused feature vector and the abnormal condition feature vector stored in the database specifically includes: calculating the cosine similarity between the fused feature vector and the abnormal condition feature vector stored in the database.

[0144] In one embodiment, the extracting features of the sensor data using a long short-term memory network to generate a first feature vector specifically includes:

[0145] Receive sensor data using an input layer of a long short-term memory network, wherein the sensor data includes a plurality of sensor parameter groups, each of the sensor parameter groups includes sensor parameters of a plurality of categories;

[0146] The input layer is used to transfer the received process parameter sequence to the convolution layer for convolution to generate the sensor data feature vector;

[0147] Flattening the sensor data feature vector;

[0148] According to the input rule of the first LSTM layer, the data shape of the flattened sensor data feature vector is adjusted;

[0149] Input the adjusted sensor data feature vector into the first LSTM layer to learn long-term dependencies and return the first hidden state sequence at each time step;

[0150] Input the first hidden state sequence into the Dropout layer, and randomly discard 25% of the data in the first hidden state sequence;

[0151] The first hidden state sequence after discarding the data is input into the second LSTM layer to extract the time-dependent features, and the sensor data sequence information corresponding to the hidden state of the last time step is returned; the sensor data sequence information is recorded as the first feature vector.

[0152] In one embodiment, the extracting features of the natural language description using the BERT model to generate a second feature vector specifically includes:

[0153] Using a word segmenter to split the natural language description to generate a word unit sequence; adding a head end marker to the head end of the word unit sequence and an end marker to the end of the word unit sequence; replacing each word unit in the word unit sequence with an index in the vocabulary to generate an index sequence;

[0154] Using the embedding layer of the BERT model, each index in the index sequence is processed respectively to generate corresponding word embedding, position embedding and tag embedding, and the sum of the word embedding, position embedding and tag embedding corresponding to each index is calculated to generate a total embedding;

[0155] The Transform encoder layer of the BERT model is used to encode each total embedding to generate a word element vector corresponding to each total embedding and a second feature vector associated with the head tag.

[0156] In one embodiment, the step of fusing the first feature vector and the second feature vector using a fully connected layer to generate a fused feature vector specifically includes:

[0157] The step of concatenating the first feature vector and the second feature vector to generate a concatenated vector includes:

[0158] concatenation vector = [first eigenvector; second eigenvector];

[0159] Among them, the dimension of the concatenated vector is n1+n2, n1 is the dimension of the first eigenvector, and n2 is the dimension of the second eigenvector;

[0160] The concatenated vector is input into the fully connected layer for fusion to generate a fused feature vector, specifically including:

[0161] Fusion feature vector = Tanh (concatenated vector W + b);

[0162] Among them, W is the weight matrix with dimension (n1+n2)×m, m is the dimension of the output of the fully connected layer; b is the bias vector with dimension m; Tanh is the activation function.

[0163] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0164] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0165] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. An intelligent inspection system, characterized in that: include: A data monitoring module, used to monitor DCS data and push the DCS data to an associated mobile terminal device; The inspection setting module is used to adjust the inspection mode according to the level of the point, including: using the verification mode for inspection when the point is a regular point; using the blind inspection mode for inspection when the point is a key point; The abnormality handling module is used to record abnormal conditions found during inspection, notify internal operators or external operators to handle them, and structure the abnormal conditions and handling operations into data; Intelligent inspection module, used to inspect the production area, including process inspection, equipment inspection and comprehensive inspection; The inspection item management module is used to summarize the inspection items in the inspection forms of each department, merge the same inspection items, and reduce repeated inspections; A record prediction module, used to record equipment inspection and maintenance records, evaluate the operating status of the equipment based on the inspection and maintenance records, and perform preventive maintenance and predictive maintenance on the equipment based on the operating status of the equipment; A model training module, used to train an abnormal condition handling model by using the abnormal condition and handling operation after structured processing as training samples; Recording and reporting module, used to enter maintenance and repair records and maintenance records into the data system; The user management module is used to manage the roles and permissions of system users.

2. An intelligent inspection method, characterized in that: Applied to the intelligent inspection system as claimed in claim 1, the intelligent inspection method specifically comprises: Using the data monitoring module to monitor DCS data, and pushing the DCS data to an associated mobile terminal device; The inspection setting module is used to adjust the inspection mode according to the level of the point, specifically including: using the verification mode for inspection when the point is a regular point; using the blind inspection mode for inspection when the point is a key point; The abnormality handling module is used to record abnormal conditions found during inspection, notify internal operators or external operators to handle them, and perform data structured processing on the abnormal conditions and handling operations; The intelligent inspection module is used to inspect the production area, including process inspection, equipment inspection and comprehensive inspection; The inspection item management module is used to summarize the inspection items in the inspection forms of each department, merge the same inspection items therein, and reduce repeated inspections; Using the record prediction module to record equipment inspection and maintenance records, evaluating the operating status of the equipment based on the inspection and maintenance records, and performing preventive maintenance and predictive maintenance on the equipment based on the operating status of the equipment; Using the model training module to take the abnormal conditions and handling operations after structured processing as training samples, training the abnormal condition handling model; Using the record reporting module to enter the maintenance and repair records and maintenance records into the data system; The user management module is used to manage the roles and permissions of system users.

3. The intelligent inspection method according to claim 2, characterized in that: The intelligent inspection method further includes: using the abnormal condition handling model to analyze the detected abnormal condition and generate a corresponding abnormal condition handling operation.

4. The intelligent inspection method according to claim 3, characterized in that: The abnormal situation handling model is used to analyze the detected abnormal situation and generate corresponding abnormal situation handling operations, specifically including: Acquire abnormal state data of the device in an abnormal condition, wherein the abnormal state data includes sensor data and a natural language description formed by an inspection personnel; Performing feature extraction on the sensor data to generate a first feature vector; Performing feature extraction on the natural language description to generate a second feature vector; Fusing the first feature vector and the second feature vector to generate a fused feature vector; Calculating the similarity between the fused feature vector and the abnormal condition feature vector stored in the database; The handling operation corresponding to the abnormal situation feature vector having the greatest similarity to the fused feature vector is used as the abnormal situation handling operation.

5. The intelligent inspection method according to claim 4, characterized in that: The extracting features from the sensor data to generate a first feature vector specifically includes: extracting features from the sensor data using a long short-term memory network to generate a first feature vector; The extracting features from the natural language description to generate a second feature vector specifically includes: extracting features from the natural language description using a BERT model to generate a second feature vector; The fusing the first feature vector and the second feature vector to generate a fused feature vector specifically includes: using a fully connected layer to fuse the first feature vector and the second feature vector to generate a fused feature vector; The calculating the similarity between the fused feature vector and the abnormal condition feature vector stored in the database specifically includes: calculating the cosine similarity between the fused feature vector and the abnormal condition feature vector stored in the database.

6. The intelligent inspection method according to claim 5, characterized in that: The extracting features of the sensor data using the long short-term memory network to generate a first feature vector specifically includes: Receive sensor data using an input layer of a long short-term memory network, wherein the sensor data includes a plurality of sensor parameter groups, each of the sensor parameter groups includes sensor parameters of a plurality of categories; The input layer is used to transfer the received process parameter sequence to the convolution layer for convolution to generate the sensor data feature vector; Flattening the sensor data feature vector; According to the input rule of the first LSTM layer, the data shape of the flattened sensor data feature vector is adjusted; Input the adjusted sensor data feature vector into the first LSTM layer to learn long-term dependencies and return the first hidden state sequence at each time step; Input the first hidden state sequence into the Dropout layer, and randomly discard 25% of the data in the first hidden state sequence; The first hidden state sequence after discarding the data is input into the second LSTM layer to extract the time-dependent features, and the sensor data sequence information corresponding to the hidden state of the last time step is returned; the sensor data sequence information is recorded as the first feature vector.

7. The intelligent inspection method according to claim 5, characterized in that: The extracting features of the natural language description using the BERT model to generate a second feature vector specifically includes: Using a word segmenter to split the natural language description to generate a word unit sequence; adding a head end marker to the head end of the word unit sequence and an end marker to the end of the word unit sequence; replacing each word unit in the word unit sequence with an index in the vocabulary to generate an index sequence; Using the embedding layer of the BERT model, each index in the index sequence is processed respectively to generate corresponding word embedding, position embedding and tag embedding, and the sum of the word embedding, position embedding and tag embedding corresponding to each index is calculated to generate a total embedding; The Transform encoder layer of the BERT model is used to encode each total embedding to generate a word element vector corresponding to each total embedding and a second feature vector associated with the head tag.

8. The intelligent inspection method according to claim 5, characterized in that: The step of fusing the first feature vector and the second feature vector using a fully connected layer to generate a fused feature vector specifically includes: The step of concatenating the first feature vector and the second feature vector to generate a concatenated vector includes: concatenation vector = [first eigenvector; second eigenvector]; Among them, the dimension of the concatenated vector is n1+n2, n1 is the dimension of the first eigenvector, and n2 is the dimension of the second eigenvector; The concatenated vector is input into the fully connected layer for fusion to generate a fused feature vector, specifically including: Fusion feature vector = Tanh (concatenated vector W + b); Among them, W is the weight matrix with dimension (n1+n2)×m, m is the dimension of the output of the fully connected layer; b is the bias vector with dimension m; Tanh is the activation function.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the intelligent inspection method as described in any one of claims 2 to 8 is implemented.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the intelligent inspection method as described in any one of claims 2 to 8 is implemented.

Images