CN117422481B

CN117422481B - State detection method and AI system for production control of fireproof high-elastic fabric

Info

Publication number: CN117422481B
Application number: CN202311736448.5A
Authority: CN
Inventors: 易亮; 曾星玮
Original assignee: Yu City Textile Technology SZ Ind Co ltd
Current assignee: Yu City Textile Technology SZ Ind Co ltd
Priority date: 2023-12-18
Filing date: 2023-12-18
Publication date: 2024-03-29
Anticipated expiration: 2043-12-18
Also published as: CN117422481A

Abstract

According to the state detection method and the AI system for the fireproof high-elastic fabric production control, in the forging risk judging process, the 'active mask' processing of the finished product quality test report to be analyzed is realized by removing the target key item quality test text, text semantic prediction vectors can be mined based on context information corresponding to the target key item quality test text in the 'active mask' finished product quality test report to be analyzed, then text semantic difference comparison is carried out by combining the authenticated text semantic vectors of the target key item quality test text before unmasked, so that whether the target key item quality test text is forged or not can be accurately and reliably judged, fault state detection is carried out on the fireproof high-elastic fabric production control system when the forging risk exists on the target key item quality test text, on one hand, the accuracy and the reliability of forging risk judgment can be improved, and on the other hand, the productivity influence caused by frequent shutdown detection on the fireproof high-elastic fabric production control system can be avoided.

Description

State detection method and AI system for production control of fireproof high-elastic fabric

Technical Field

The application relates to the technical field of artificial intelligence, in particular to a state detection method and an AI system for production control of fireproof high-elastic fabrics.

Background

Fireproof high-elastic fabric (also called as high-elastic flame-retardant fabric) uses all-cotton fabric, polyester-cotton fabric, cotton-nylon blended fabric and the like as base fabric, and the fabric is used for protective clothing, labor protection clothing, working clothing materials and the like. The fireproof high-elastic fabric is mainly used in the inflammable and explosive high-risk fields of fire-fighting clothing, deep forest rescue clothing, electric welder labor protection clothing, railways, aviation, electric power, petroleum and the like, and is particularly important for production control quality control of the fireproof high-elastic fabric.

Disclosure of Invention

In order to improve the technical problems in the related art, the application provides a state detection method and an AI system for production control of fireproof high-elastic fabrics.

In a first aspect, an embodiment of the present application provides a method for detecting a state of production control of a fireproof high-elastic fabric, which is applied to an AI system, and the method includes:

acquiring a to-be-analyzed fireproof high-elastic fabric quality inspection text set corresponding to a to-be-analyzed finished product quality test report, and acquiring an authenticated text semantic vector of a target key item quality inspection text corresponding to the to-be-analyzed fireproof high-elastic fabric quality inspection text set, wherein the to-be-analyzed fireproof high-elastic fabric quality inspection text set is obtained by removing the target key item quality inspection text from the target fireproof high-elastic fabric quality inspection text set corresponding to the to-be-analyzed finished product quality test report;

Loading the to-be-analyzed fireproof high-elastic fabric quality inspection text set into a target production control state analysis network to obtain text semantic prediction vectors corresponding to the target key item quality inspection text;

determining a report forging discrimination tag corresponding to the quality test report of the finished product to be analyzed according to the text semantic difference between the authenticated text semantic vector and the text semantic prediction vector;

and when the report forging discrimination tag characterizes that the to-be-analyzed finished product quality test report has forging risk, detecting a fault state of a fireproof high-elastic fabric production control system corresponding to the to-be-analyzed finished product quality test report.

In some optional examples, the method for debugging the target production control state analysis network includes:

acquiring a fireproof high-elastic fabric quality inspection text sample set and text semantic vector samples corresponding to the fireproof high-elastic fabric quality inspection text sample set, and acquiring target text semantic vectors of key item quality inspection text samples corresponding to the fireproof high-elastic fabric quality inspection text sample set, wherein the fireproof high-elastic fabric quality inspection text sample set is obtained by removing the key item quality inspection text samples from an original fireproof high-elastic fabric quality inspection text set;

Loading the fireproof high-elastic fabric quality inspection text sample set to an original production control state analysis network to obtain text semantic recognition vectors corresponding to the key item quality inspection text samples;

generating a flow positive adjustment sample of a flow decision component corresponding to the original production control state analysis network according to the target text semantic vector and the text semantic vector sample, and generating a flow negative adjustment sample corresponding to the flow decision component according to the text semantic vector sample and the text semantic recognition vector;

according to the target text semantic vector, the text semantic recognition vector, the flow positive debugging sample and the flow negative debugging sample, the original production control state analysis network and the flow decision component are subjected to joint debugging to obtain joint debugging cost;

and optimizing the network variables of the original production control state analysis network and the stream decision component according to the joint debugging cost until the network variables meet the debugging termination requirement, and obtaining a target production control state analysis network.

In some optional examples, the obtaining a set of text semantic vector samples corresponding to the set of text samples for fire-resistant high-elastic fabric quality inspection includes:

Acquiring a finished product quality test report sample, and acquiring a plurality of standby fireproof high-elastic fabric quality inspection texts from the finished product quality test report sample, wherein a report forging discrimination label corresponding to the finished product quality test report sample is that report forging does not exist;

detecting Wen Naila quality inspection event resistance of each standby fireproof high-elastic fabric quality inspection text to obtain Wen Naila quality inspection event text blocks corresponding to each standby fireproof high-elastic fabric quality inspection text;

obtaining a Wen Naila quality inspection event text corresponding to each standby fireproof high-elastic fabric quality inspection text through each Wen Naila quality inspection event text block, and obtaining an original fireproof high-elastic fabric quality inspection text set through each Wen Naila quality inspection event text;

extracting the fireproof high-elastic fabric quality inspection text from the original fireproof high-elastic fabric quality inspection text set to obtain a fireproof high-elastic fabric quality inspection text sample set;

and carrying out text semantic vector mining on the fireproof high-elastic fabric quality inspection text sample set to obtain the text semantic vector sample.

In some optional examples, the text semantic vector examples include text semantic vector examples corresponding to each fire-proof high-elastic fabric quality inspection text example in the fire-proof high-elastic fabric quality inspection text example set, the text semantic vector examples, the target text semantic vector and the text semantic recognition vector all include text semantic features of at least one description layer, and the key item quality inspection text example and the fire-proof high-elastic fabric quality inspection text example all add a quality inspection text digital authentication signature;

The generating a positive flow adjustment sample of the flow decision component corresponding to the original production control state analysis network according to the target text semantic vector and the text semantic vector sample, and generating a negative flow adjustment sample corresponding to the flow decision component according to the text semantic vector sample and the text semantic recognition vector, comprising:

determining text semantic stream characteristics according to the quality inspection text sample of the key item and the quality inspection text digital authentication signature corresponding to the quality inspection text sample of each fireproof high-elastic fabric;

according to the text semantic flow characteristics, the text semantic characteristics of the same description level in the target text semantic vector and the text semantic vector sample are aggregated to obtain flow positive adjustment samples corresponding to each description level;

and according to the text semantic flow characteristics, the text semantic recognition vectors and the text semantic characteristics of the same description level in the text semantic vector samples are aggregated to obtain flow type negative debugging samples corresponding to each description level.

In some optional examples, the performing joint debugging on the original production control state analysis network and the streaming decision component according to the target text semantic vector, the text semantic recognition vector, the streaming positive debugging sample and the streaming negative debugging sample to obtain joint debugging cost includes:

Generating text semantic cost according to the target text semantic vector and the text semantic recognition vector;

loading the flow positive adjustment sample and the flow negative adjustment sample into the flow decision component to obtain a first identification viewpoint corresponding to the flow positive adjustment sample and a second identification viewpoint corresponding to the flow negative adjustment sample;

generating a streaming cost according to a first recognition viewpoint and a first priori viewpoint corresponding to the streaming positive adjustment sample, and a second recognition viewpoint and a second priori viewpoint corresponding to the streaming negative adjustment sample;

and generating the joint debugging cost according to the text semantic cost and the streaming cost.

In some optional examples, the target text semantic vector includes target text semantic features corresponding to at least two description levels, and the text semantic recognition vector includes text semantic feature prediction results corresponding to the at least two description levels;

the generating text semantic costs according to the target text semantic vector and the text semantic recognition vector comprises the following steps:

obtaining a first debugging cost based on the difference between the target text semantic feature corresponding to the authenticated description level and the text semantic feature prediction result from the target text semantic vector and the text semantic recognition vector;

Performing semantic adjustment processing on the target text semantic features corresponding to the authenticated description level to obtain corresponding first text semantic adjustment vectors, and performing semantic adjustment processing on text semantic feature prediction results corresponding to the authenticated description level to obtain corresponding second text semantic adjustment vectors;

obtaining a second debugging cost according to the difference between the first text semantic adjustment vector and the second text semantic adjustment vector;

obtaining a third debugging cost based on the difference between the target text semantic feature and the text semantic feature prediction result of the residual description level from the target text semantic vector and the text semantic recognition vector;

and obtaining the text semantic cost according to the first debugging cost, the second debugging cost and the third debugging cost.

In some optional examples, the current text semantic feature is a target text semantic feature or a text semantic feature prediction result corresponding to the authenticated description layer, and the semantic adjustment processing is performed on the current text semantic feature to obtain a corresponding current text semantic adjustment vector, including:

downsampling the current text semantic features to obtain first text semantic features;

Upsampling the first text semantic features to obtain second text semantic features;

the second text semantic features and the current text semantic features have consistent vector dimensions;

obtaining a target semantic difference based on the difference between the current text semantic feature and the second text semantic feature;

taking the first text semantic feature as an adjusted current text semantic feature, and jumping to the step of downsampling the current text semantic feature until the first text semantic feature meets debugging expectations, so as to obtain a plurality of target semantic differences with queue priorities;

and obtaining the current text semantic adjustment vector through each target semantic difference.

In some optional examples, the streaming decision component includes at least one streaming analysis component corresponding to a description level, the streaming positive adjustment sample includes a streaming positive adjustment sample corresponding to each description level, and the streaming negative adjustment sample includes a streaming negative adjustment sample corresponding to each description level;

the loading the flow positive adjustment sample and the flow negative adjustment sample into the flow decision component to obtain a first recognition viewpoint corresponding to the flow positive adjustment sample and a second recognition viewpoint corresponding to the flow negative adjustment sample, including: and respectively loading the streaming positive adjustment sample and the streaming negative adjustment sample of the same description level to corresponding streaming analysis components to obtain a first recognition viewpoint and a second recognition viewpoint corresponding to each description level.

In some alternative examples, the method further comprises:

generating channel attention active adjustment sample cases of a channel branch network corresponding to the original production control state analysis network according to the target text semantic vector, and generating channel attention active adjustment sample cases corresponding to the channel branch network according to the target text semantic vector and the text semantic recognition vector;

according to the target text semantic vector, the text semantic recognition vector, the channel attention active adjustment sample and the channel attention passive adjustment sample, carrying out joint debugging on the original production control state analysis network and the channel branch network to obtain updated debugging cost;

and optimizing the network variables of the original production control state analysis network and the channel branch network according to the updating and debugging cost until the network variables meet the debugging termination requirement, and obtaining the target production control state analysis network.

In some alternative examples, the method further comprises:

according to the target text semantic vector, the text semantic recognition vector, the channel attention positive adjustment sample, the channel attention negative adjustment sample, the flow positive adjustment sample and the flow negative adjustment sample, carrying out joint debugging on the original production control state analysis network, the channel branch network and the flow decision component to obtain target debugging cost;

And optimizing network variables of the original production control state analysis network, the channel branch network and the flow decision component according to the target debugging cost until the network variables meet the debugging termination requirement, so as to obtain the target production control state analysis network.

In some optional examples, the target text semantic vector includes a minimum of two description levels of target text semantic features, and the text semantic recognition vector includes a minimum of two description levels of text semantic feature predictors;

the generating the channel attention negative test sample corresponding to the channel branch network according to the target text semantic vector and the text semantic recognition vector comprises the following steps: and generating the channel attention negative adjustment sample according to the target text semantic features of the target description level in the target text semantic vector and the text semantic feature prediction results of the residual description levels in the text semantic recognition vector.

In a second aspect, the present application also provides an AI system, including a processor and a memory; the processor is in communication with the memory, and the processor is configured to read and execute a computer program from the memory to implement the method described above.

In a third aspect, the present application also provides a computer readable storage medium having stored thereon a program which, when executed by a processor, implements the method described above.

By applying the embodiment of the application, in the forging risk judging process, the 'active mask' processing of the finished product quality test report to be analyzed is realized by removing the target key item quality test text, text semantic prediction vectors can be mined based on the context information corresponding to the target key item quality test text in the 'active mask' finished product quality test report to be analyzed, then text semantic difference comparison is carried out by combining the authenticated text semantic vectors of the target key item quality test text before being not masked, so that whether the target key item quality test text is forged or not can be accurately and reliably judged, fault state detection is carried out on the fireproof high-elastic fabric production control system when the forging risk exists in the target key item quality test text, on one hand, the accuracy and the reliability of the forging risk judgment can be improved, and on the other hand, the influence on the productivity caused by frequent shutdown detection on the fireproof high-elastic fabric production control system can be avoided.

Further, a fireproof high-elastic fabric quality inspection text sample set and text semantic vector samples corresponding to the fireproof high-elastic fabric quality inspection text sample set are obtained, target text semantic vectors of key item quality inspection text samples corresponding to the fireproof high-elastic fabric quality inspection text sample set are obtained by removing the key item quality inspection text samples from an original fireproof high-elastic fabric quality inspection text set, and the fireproof high-elastic fabric quality inspection text sample set is loaded to an original production control state analysis network to obtain text semantic recognition vectors corresponding to the key item quality inspection text samples. In this way, the quality inspection text samples of the key items are removed in the fireproof high-elastic fabric quality inspection text sample set, namely, the fireproof high-elastic fabric quality inspection text sample set is a report set with a mask, and the fireproof high-elastic fabric quality inspection text sample set is loaded to a production control state analysis network to analyze text semantic vectors of the quality inspection text samples of the key items. Generating a flow positive adjustment sample of a flow decision component corresponding to the original production control state analysis network based on the target text semantic vector and the text semantic vector sample, generating a flow negative adjustment sample corresponding to the flow decision component based on the text semantic vector sample and the text semantic recognition vector, and performing joint debugging on the original production control state analysis network and the flow decision component based on the target text semantic vector, the text semantic recognition vector, the flow positive adjustment sample and the flow negative adjustment sample to obtain joint debugging cost, optimizing network variables of the original production control state analysis network and the flow decision component based on the joint debugging cost until the network variables meet debugging termination requirements, thereby obtaining the target production control state analysis network. In this way, the target text semantic vector and the text semantic vector sample are accurate and correct text semantic vectors, the text semantic recognition vector is a text semantic vector obtained through AI recognition, a positive sample is generated through the accurate and correct text semantic vector, a negative sample is generated through the accurate and correct text semantic vector and the recognized text semantic vector, and the production control state analysis network and the streaming decision component are subjected to joint debugging through the target text semantic vector, the text semantic recognition vector, the positive sample and the negative sample to realize the debugging of the AI neural network, so that the recognition performance of the production control state analysis network on a time sequence level can be improved based on the streaming decision component, and further the production control state analysis network with high report forging discrimination precision is debugged.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and together with the description, serve to explain the principles of the application.

Fig. 1 is a schematic flow chart of a state detection method for production control of fireproof high-elastic fabrics according to an embodiment of the present application.

Detailed Description

Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numbers in different drawings refer to the same or similar elements, unless otherwise indicated. The implementations described in the following exemplary examples are not representative of all implementations consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present application as detailed in the accompanying claims.

It should be noted that the terms "first," "second," and the like in the description and claims of the present application and the above figures are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order.

The method embodiments provided by the embodiments of the present application may be performed in an AI system, a computer device, or a similar computing apparatus. Taking the example of operation on an AI system, the AI system may include one or more processors (which may include, but are not limited to, a microprocessor MCU, a processing device such as a programmable logic device FPGA) and a memory for storing data, and optionally, a transmission device for communication functions. It will be appreciated by those of ordinary skill in the art that the above-described structure is merely illustrative and is not intended to limit the structure of the above-described AI system. For example, the AI system may also include more or fewer components than shown above, or have a different configuration than shown above.

The memory may be used to store a computer program, for example, a software program of application software and a module, for example, a computer program corresponding to a method for detecting a state of production control of fireproof high-elastic fabric in the embodiments of the present application, and the processor executes the computer program stored in the memory, thereby performing various functional applications and data processing, that is, implementing the method described above. The memory may include high speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid state memory. In some examples, the memory may further include memory remotely located with respect to the processor, which may be connected to the AI system via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The transmission means is used for receiving or transmitting data via a network. The specific example of the network described above may include a wireless network provided by a communication provider of an AI system. In one example, the transmission means comprises a network adapter (Network Interface Controller, simply referred to as NIC) that can be connected to other network devices via a base station to communicate with the internet. In one example, the transmission device may be a Radio Frequency (RF) module, which is used to communicate with the internet wirelessly.

Referring to fig. 1, fig. 1 is a flow chart of a state detection method for production control of fireproof high-elastic fabric according to an embodiment of the present application, where the method is applied to an AI system, and further includes steps 102-108.

And 102, acquiring a quality inspection text set of the fireproof high-elastic fabric to be analyzed corresponding to the quality test report of the finished product to be analyzed, and acquiring a verified text semantic vector of a quality inspection text of a target key item corresponding to the quality inspection text set of the fireproof high-elastic fabric to be analyzed.

Further, the finished product quality test report to be analyzed refers to a text report to be analyzed for the presence of falsification of quality inspection results. The target fireproof high-elastic fabric quality inspection text set refers to a fireproof high-elastic fabric quality inspection text set corresponding to the to-be-analyzed finished product quality test report, and the fireproof high-elastic fabric quality inspection text included in the target fireproof high-elastic fabric quality inspection text set is obtained from the to-be-analyzed finished product quality test report.

Further, the to-be-analyzed fireproof high-elastic fabric quality inspection text set refers to a fireproof high-elastic fabric quality inspection text set with a mask to be input into a target production control state analysis network, and is used for determining whether a quality inspection result of a to-be-analyzed finished product quality test report is forged or not. The to-be-analyzed fireproof high-elastic fabric quality inspection text set is obtained by removing target key item quality inspection texts from target fireproof high-elastic fabric quality inspection text sets corresponding to-be-analyzed finished product quality test reports. The target key item quality inspection text refers to key item quality inspection text corresponding to the fireproof high-elastic fabric quality inspection text set to be analyzed.

Under the application scene of the application, the special use environment of the fireproof high-elastic fabric is very important for quality detection analysis and authenticity discrimination of the quality detection analysis of the fireproof high-elastic fabric. The inventor finds that the traditional quality detection analysis processing only aims at surface analysis of a quality inspection report, and does not judge authenticity of quality inspection texts of some key items (such as high-temperature-resistant quality inspection items, pull-resistant quality inspection items and the like), so that if quality inspection data information in the quality inspection texts of the key items is forged, the quality inspection data information can have great influence on subsequent fireproof high-elastic fabric quality use. Therefore, the method and the device can remove the target key item quality inspection text based on the 'active mask' mode, predict the predicted feature of the target key item quality inspection text through the context, then compare the predicted feature with the initial feature, so as to judge whether the key item quality inspection text has a forging risk, and further detect the fault state of the fireproof high-elastic fabric production control system according to the judging result.

In the embodiment of the application, the authenticated text semantic vector refers to a priori text semantic vector and a correct text semantic vector corresponding to the quality inspection text of the target key item. The authenticated text semantic vector may include at least one text semantic feature that describes a level (feature dimension).

For example, the AI system may obtain, locally or from other systems, a set of quality inspection texts of the fire-proof high-elastic fabric to be analyzed corresponding to the quality test report of the finished product to be analyzed, and obtain a verified text semantic vector of a quality inspection text of a target key item corresponding to the set of quality inspection texts of the fire-proof high-elastic fabric to be analyzed. For example, the AI system acquires a quality test record sent by the test server as a quality test report of a finished product to be analyzed, extracts a target fireproof high-elastic fabric quality test text set from the quality test report of the finished product to be analyzed, extracts the fireproof high-elastic fabric quality test text from the target fireproof high-elastic fabric quality test text set, uses the extracted target fireproof high-elastic fabric quality test text set as the quality test text set of the fireproof high-elastic fabric to be analyzed, and uses the extracted fireproof high-elastic fabric quality test text as a target key item quality test text corresponding to the quality test text set of the fireproof high-elastic fabric to be analyzed. Further, the AI system performs text semantic mining on the target key item quality inspection text to obtain an authenticated text semantic vector corresponding to the target key item quality inspection text.

In some examples, in order to improve the report forgery discrimination precision, when the target fireproof high-elastic fabric quality inspection text set is generated based on the to-be-analyzed finished product quality test report, different combination situations of the fireproof high-elastic fabric quality inspection texts in the to-be-analyzed finished product quality test report can be considered as far as possible, and the fireproof high-elastic fabric quality inspection texts in the target fireproof high-elastic fabric quality inspection text set are the fireproof high-elastic fabric quality inspection texts adjacent in time sequence. For example, the quality test report of the finished product to be analyzed comprises fireproof high-elastic fabric quality test texts 1-text10, 1 target fireproof high-elastic fabric quality test text set is formed by 5 fireproof high-elastic fabric quality test texts with adjacent time sequences, and 6 target fireproof high-elastic fabric quality test text sets can be respectively fireproof high-elastic fabric quality test texts 1-text5, fireproof high-elastic fabric quality test texts 2-text6, fireproof high-elastic fabric quality test texts 3-text7, fireproof high-elastic fabric quality test texts 4-text8, fireproof high-elastic fabric quality test texts 5-text9 and fireproof high-elastic fabric quality test texts 6-text10. If a set of quality inspection texts of the fireproof high-elastic fabric to be analyzed corresponds to a quality inspection text of a target key item, all possible situations can be considered. For example, the target fireproof high-elastic fabric quality inspection text set comprises fireproof high-elastic fabric quality inspection text1-text5, so 5 to-be-analyzed fireproof high-elastic fabric quality inspection text sets can be obtained based on the target fireproof high-elastic fabric quality inspection text set, wherein the fireproof high-elastic fabric quality inspection text sets are respectively fireproof high-elastic fabric quality inspection text2-text5, fireproof high-elastic fabric quality inspection text1-text3-text4-text5, fireproof high-elastic fabric quality inspection text1-text2-text3-text5 and fireproof high-elastic fabric quality inspection text1-text4. It can be appreciated that if the report forgery discrimination tag obtained by any one of the target fire-resistant high-elastic fabric quality inspection text sets is the existence report forgery, the quality test report of the finished product to be analyzed is determined to be the existence report forgery.

And 104, loading the to-be-analyzed fireproof high-elastic fabric quality inspection text set into a target production control state analysis network to obtain a text semantic prediction vector corresponding to the target key item quality inspection text.

The target production control state analysis network is a debugged production control state analysis network. The text semantic prediction vector refers to an identified text semantic vector corresponding to the quality inspection text of the target key item, and is a text semantic vector obtained by reasoning based on all fireproof high-elastic fabric quality inspection texts except the quality inspection text of the target key item (namely, the context information of the quality inspection text of the target key item).

The AI system can load the to-be-analyzed fireproof high-elastic fabric quality inspection text set to a target production control state analysis network, and the target production control state analysis network performs semantic feature extraction on the target fireproof high-elastic fabric quality inspection text set so as to output a text semantic prediction vector corresponding to the target key item quality inspection text.

And 106, determining a report forgery discrimination tag corresponding to the finished product quality test report to be analyzed based on the text semantic difference between the authenticated text semantic vector and the text semantic prediction vector.

The text semantic difference refers to cosine distance between the authenticated text semantic vector and the text semantic prediction vector.

For example, the AI system may determine a report forgery decision tag corresponding to the finished product quality test report to be analyzed based on the text semantic distinction between the authenticated text semantic vector and the text semantic prediction vector. Report forgery discrimination tags include the presence of report forgery and the absence of report forgery. If the text semantic difference is large, the report forging discrimination label is that report forging exists, and if the text semantic difference is small, the report forging discrimination label is that report forging does not exist. Further, the size of the text semantic difference may be determined according to a preset text semantic difference discrimination value. For example, if the text semantic difference is higher than the text semantic difference discrimination value, the report forging discrimination label is reporting forging, and if the text semantic difference is smaller than or equal to the text semantic difference discrimination value, the report forging discrimination label is reporting forging.

And step 108, when the report forging discrimination tag represents that the to-be-analyzed finished product quality test report has forging risk, detecting a fault state of a fireproof high-elastic fabric production control system corresponding to the to-be-analyzed finished product quality test report.

In the embodiment of the application, if the text semantic difference is large, the report forgery judgment label is that report forgery exists. On the basis, the fact that the quality inspection text of the target key item corresponding to the quality test report of the finished product to be analyzed possibly fails to reach the standard is indicated, so that the failure state detection can be carried out on the fireproof high-elastic fabric production control system corresponding to the quality test report of the finished product to be analyzed, the trace source of the failure of the quality inspection text of the target key item is realized, and after the links with failures in the fireproof high-elastic fabric production control system are detected, the corresponding failure links can be maintained, and the yield of subsequent production is guaranteed.

It can be seen that, in the process of discriminating the risk of forging, the steps 102-108 are applied, the "active mask" process of the quality test report of the finished product to be analyzed is implemented by removing the quality test text of the target key item, the text semantic prediction vector can be mined based on the context information corresponding to the quality test text of the target key item in the quality test report of the finished product to be analyzed of the "active mask", and then the text semantic discrimination comparison is performed by combining the authenticated text semantic vector of the quality test text of the target key item before the mask is not covered, so that whether the quality test text of the target key item is forged or not can be accurately and reliably discriminated, and the fault state detection is performed on the fire-proof high-elastic fabric production control system when the quality test text of the target key item is forged, so that on one hand, the accuracy and the reliability of discriminating the risk of forging can be improved, and on the other hand, the productivity influence caused by frequent shutdown detection on the fire-proof high-elastic fabric production control system can be avoided.

In some independent embodiments, the fault state detection of the fireproof high-elastic fabric production control system corresponding to the to-be-analyzed finished product quality test report comprises: acquiring an infrared flaw detection image set of the fireproof high-elastic fabric production control system; and carrying out structural fault detection on each group of infrared flaw detection images in the infrared flaw detection image set to obtain a structural fault detection result.

In the embodiment of the application, the infrared flaw detection image set can be obtained by adopting infrared flaw detection equipment to acquire images of different production equipment of the fireproof high-elastic fabric production control system. Each group of infrared flaw detection images in the infrared flaw detection image set corresponds to different production equipment of the fireproof high-elastic fabric production control system. Therefore, structural fault detection is carried out on each group of infrared flaw detection images, and production equipment does not need to be disassembled in advance, so that the efficiency of structural fault detection is improved.

In some independent embodiments, the performing structural fault detection on each group of infrared flaw detection images in the infrared flaw detection image set to obtain a structural fault detection result includes: performing wavelet description mining on each group of infrared flaw detection images to obtain at least one wavelet descriptor to be detected and at least one associated wavelet descriptor corresponding to each wavelet descriptor to be detected of the at least one wavelet descriptor to be detected; pairing the at least one wavelet descriptor to be detected with preset wavelet descriptors of each fault label to obtain an initial fault label corresponding to each wavelet descriptor to be detected; determining a certainty parameter of each associated wavelet descriptor in the at least one associated wavelet descriptor according to the set wavelet description set of each fault label and the initial fault label aiming at each wavelet descriptor to be detected; the set wavelet description set contains countermeasure descriptors; determining the fault detection confidence of each wavelet descriptor to be detected based on the certainty parameter of each associated wavelet descriptor; and determining fault labels of each group of infrared flaw detection images based on the fault detection confidence coefficient of each wavelet descriptor to be detected.

It can be understood that, by applying the above embodiment, the wavelet description sub-layer mining analysis can be performed on each group of infrared flaw detection images, so that the fault detection confidence coefficient is further determined according to the obtained initial fault label, and thus, when determining the fault label of each group of infrared flaw detection images, the fault label is assisted by the set wavelet description set containing the counterdescriptor, thereby ensuring the accuracy and reliability of the fault label and avoiding detection errors.

In some alternative embodiments, the method of commissioning the target production control state analysis network includes steps 201-205.

Step 201, a fireproof high-elastic fabric quality inspection text sample set and text semantic vector samples corresponding to the fireproof high-elastic fabric quality inspection text sample set are obtained, target text semantic vectors of key item quality inspection text samples corresponding to the fireproof high-elastic fabric quality inspection text sample set are obtained, and the fireproof high-elastic fabric quality inspection text sample set is obtained by removing the key item quality inspection text samples from an original fireproof high-elastic fabric quality inspection text set.

In the embodiments of the present application, the above-mentioned examples may be understood as examples or examples for performing training of a neural network. The text semantic vector sample can be understood as the characteristic data of the sample sequence, and the target text semantic vector is the characteristic data of the key item quality inspection text sample before unmasked.

Step 202, loading the fireproof high-elastic fabric quality inspection text sample set to an original production control state analysis network to obtain text semantic recognition vectors corresponding to the key item quality inspection text samples.

The text semantic recognition vector is obtained by carrying out semantic prediction on a fireproof high-elastic fabric quality inspection text sample set after masking of the key item quality inspection text sample is completed, and is used for representing prediction feature data of the key item quality inspection text sample, and the text semantic recognition vector is obtained based on context information of the key item quality inspection text sample.

Step 203, generating a positive flow adjustment sample of the flow decision component corresponding to the original production control state analysis network according to the target text semantic vector and the text semantic vector sample, and generating a negative flow adjustment sample corresponding to the flow decision component according to the text semantic vector sample and the text semantic recognition vector.

In this embodiment, the flow decision component is a time sequence analysis subnet based on a time sequence feature layer structure, based on which, the flow positive adjustment sample can be understood as a positive sample based on the time sequence feature layer, and the flow negative adjustment sample is a negative sample based on the time sequence feature layer.

And 204, performing joint debugging on the original production control state analysis network and the flow decision component according to the target text semantic vector, the text semantic recognition vector, the flow positive debugging sample and the flow negative debugging sample to obtain joint debugging cost.

In this embodiment of the present application, the joint debugging may be countermeasure debugging, for example, performing countermeasure debugging on the original production control state analysis network and the streaming decision component respectively in combination with the target text semantic vector, the text semantic recognition vector, the streaming positive debugging sample and the streaming negative debugging sample, so as to obtain a loss function after the countermeasure debugging.

And step 205, optimizing the network variables of the original production control state analysis network and the flow decision component according to the joint debugging cost until the network variables meet the debugging termination requirement, and obtaining a target production control state analysis network.

In the embodiment of the application, the network variables of the original production control state analysis network and the streaming decision component can be improved and optimized towards the convergence direction of the joint debugging cost until the requirement of debugging termination is met (for example, the joint debugging cost converges, or the iterative debugging times reach the set times, etc.).

And (3) applying steps 201-205 to obtain a fireproof high-elastic fabric quality inspection text sample set and text semantic vector samples corresponding to the fireproof high-elastic fabric quality inspection text sample set, obtaining target text semantic vectors of key item quality inspection text samples corresponding to the fireproof high-elastic fabric quality inspection text sample set, wherein the fireproof high-elastic fabric quality inspection text sample set is obtained by removing the key item quality inspection text samples from an original fireproof high-elastic fabric quality inspection text set, and loading the fireproof high-elastic fabric quality inspection text sample set to an original production control state analysis network to obtain text semantic recognition vectors corresponding to the key item quality inspection text samples. In this way, the quality inspection text samples of the key items are removed in the fireproof high-elastic fabric quality inspection text sample set, namely, the fireproof high-elastic fabric quality inspection text sample set is a report set with a mask, and the fireproof high-elastic fabric quality inspection text sample set is loaded to a production control state analysis network to analyze text semantic vectors of the quality inspection text samples of the key items. Generating a flow positive adjustment sample of a flow decision component corresponding to the original production control state analysis network based on the target text semantic vector and the text semantic vector sample, generating a flow negative adjustment sample corresponding to the flow decision component based on the text semantic vector sample and the text semantic recognition vector, and performing joint debugging on the original production control state analysis network and the flow decision component based on the target text semantic vector, the text semantic recognition vector, the flow positive adjustment sample and the flow negative adjustment sample to obtain joint debugging cost, optimizing network variables of the original production control state analysis network and the flow decision component based on the joint debugging cost until the network variables meet debugging termination requirements, thereby obtaining the target production control state analysis network. In this way, the target text semantic vector and the text semantic vector sample are accurate and correct text semantic vectors, the text semantic recognition vector is a text semantic vector obtained through AI recognition, a positive sample is generated through the accurate and correct text semantic vector, a negative sample is generated through the accurate and correct text semantic vector and the recognized text semantic vector, and the production control state analysis network and the streaming decision component are subjected to joint debugging through the target text semantic vector, the text semantic recognition vector, the positive sample and the negative sample to realize the debugging of the AI neural network, so that the recognition performance of the production control state analysis network on a time sequence level can be improved based on the streaming decision component, and further the production control state analysis network with high report forging discrimination precision is debugged.

In some possible embodiments, the step 201 of obtaining a set of text samples for quality inspection of fireproof high-elastic fabric and text semantic vector samples corresponding to the set of text samples for quality inspection of fireproof high-elastic fabric includes steps 2011-2015.

And 2011, acquiring a finished product quality test report sample, and acquiring a plurality of standby fireproof high-elastic fabric quality inspection texts from the finished product quality test report sample, wherein a report forging judgment label corresponding to the finished product quality test report sample is that report forging does not exist.

And 2012, detecting Wen Naila quality inspection event resistance of each standby fireproof high-elastic fabric quality inspection text to obtain Wen Naila quality inspection event resistance text blocks corresponding to each standby fireproof high-elastic fabric quality inspection text.

The Wen Naila quality inspection event text block can record high temperature detection data (such as text data or chart data) for the fireproof high-elastic fabric and pulling deformation detection data (such as text data or chart data) for the fireproof high-elastic fabric.

And 2013, obtaining a Wen Naila quality inspection event text corresponding to the quality inspection text of each standby fireproof high-elasticity fabric through each Wen Naila quality inspection event text block, and obtaining an original fireproof high-elasticity fabric quality inspection text set through each Wen Naila quality inspection event text.

2014, extracting the fireproof high-elasticity fabric quality inspection text of the original fireproof high-elasticity fabric quality inspection text set to obtain the fireproof high-elasticity fabric quality inspection text sample set;

and step 2015, performing text semantic vector mining on the fireproof high-elastic fabric quality inspection text sample set to obtain the text semantic vector sample.

According to the embodiment of the application, the Wen Naila quality inspection event text block corresponding to the standby fireproof high-elastic fabric quality inspection text can be accurately segmented through detecting the Wen Naila quality inspection event, so that the original fireproof high-elastic fabric quality inspection text set and the subsequent fireproof high-elastic fabric quality inspection text sample set are accurately obtained based on the Wen Naila quality inspection event text block, and therefore high-quality and complete expression of the obtained text semantic vector sample on the detail content of the Wen Naila quality inspection event can be ensured when text semantic vector mining is carried out.

In some possible embodiments, the text semantic vector samples include text semantic vector samples corresponding to each fire-resistant high-elastic fabric quality inspection text sample in the set of fire-resistant high-elastic fabric quality inspection text samples. Further, the text semantic vector sample, the target text semantic vector, and the text semantic recognition vector all contain at least one descriptive-level text semantic feature. Furthermore, the quality inspection text sample of the key item and the quality inspection text sample of the fireproof high-elastic fabric are added with quality inspection text digital authentication signatures. In the embodiment of the application, the quality inspection text digital authentication signature can be understood as text timestamp information corresponding to the quality inspection text sample of the key item and the quality inspection text sample of the fireproof high-elastic fabric. Based on this, the generating a positive flow adjustment sample of the flow decision component corresponding to the original production control state analysis network according to the target text semantic vector and the text semantic vector sample in step 203 generates a negative flow adjustment sample corresponding to the flow decision component according to the text semantic vector sample and the text semantic recognition vector, which includes steps 2031-2033.

Step 2031, determining text semantic stream characteristics according to the quality inspection text sample of the key item and the quality inspection text digital authentication signature corresponding to the quality inspection text sample of each fireproof high-elastic fabric.

In the embodiment of the application, text semantic streaming features can be understood as time sequence feature data of text semantic vectors.

Step 2032, aggregating the text semantic features of the same description level in the target text semantic vector and the text semantic vector sample according to the text semantic streaming features to obtain streaming positive adjustment samples corresponding to each description level.

Step 2033, aggregating the text semantic recognition vector and the text semantic features of the same description level in the text semantic vector sample according to the text semantic streaming features to obtain streaming negative adjustment samples corresponding to each description level.

In the embodiment of the application, text semantic stream type feature mining is performed based on the quality inspection text digital authentication signature carried by the corresponding text sample, and the stream type positive test sample and the stream type negative test sample can be accurately and completely obtained through feature aggregation processing.

In some optional embodiments, the step 204 includes performing joint debugging on the raw production control state analysis network and the streaming decision component according to the target text semantic vector, the text semantic recognition vector, the streaming positive debugging sample and the streaming negative debugging sample to obtain a joint debugging cost, which includes steps 2041-2044.

Step 2041, generating text semantic cost according to the target text semantic vector and the text semantic recognition vector.

In the embodiment of the application, the text semantic cost can be understood as a loss function generated from the text semantic feature mining perspective.

Step 2042, loading the streaming positive adjustment sample and the streaming negative adjustment sample into the streaming decision component to obtain a first recognition viewpoint corresponding to the streaming positive adjustment sample and a second recognition viewpoint corresponding to the streaming negative adjustment sample.

In this embodiment of the present application, the first recognition viewpoint may be understood as a positive case prediction tag, and the second recognition viewpoint may be understood as a negative case prediction tag.

Step 2043, generating a streaming cost according to the first recognition viewpoint and the first priori viewpoint corresponding to the streaming positive adjustment sample, and the second recognition viewpoint and the second priori viewpoint corresponding to the streaming negative adjustment sample.

Still further, the first prior view may understand the positive case realistic tag and the second prior view may understand the negative case realistic tag. On the basis, the loss function of the time sequence characteristic layer can be accurately generated by integrating the recognition viewpoints/prior viewpoints of the positive and negative samples.

And 2044, generating the joint debugging cost according to the text semantic cost and the streaming cost.

On the basis of the above, the text semantic cost and the stream cost can be integrated to determine the joint debugging cost, so that the loss function attention to the text semantic dimension and the time sequence dimension is realized, and the accuracy of the joint debugging cost is improved.

In some examples, the target text semantic vector includes target text semantic features corresponding to a minimum of two description levels, and the text semantic recognition vector includes text semantic feature predictors (text feature predictors) corresponding to the minimum of two description levels. Based on this, the generating text semantic costs according to the target text semantic vector and the text semantic recognition vector in step 2041 includes steps 20411-20415.

Step 20411, obtaining a first debugging cost from the target text semantic vector and the text semantic recognition vector based on the difference between the target text semantic feature corresponding to the authenticated description level and the text semantic feature prediction result.

Step 20412, performing semantic adjustment processing on the target text semantic features corresponding to the authenticated description level to obtain corresponding first text semantic adjustment vectors, and performing semantic adjustment processing on the text semantic feature prediction results corresponding to the authenticated description level to obtain corresponding second text semantic adjustment vectors.

Step 20413, obtaining a second debugging cost according to the difference between the first text semantic adjustment vector and the second text semantic adjustment vector.

Step 20414, obtaining a third debugging cost based on the difference between the target text semantic feature and the text semantic feature prediction result of the residual description level from the target text semantic vector and the text semantic recognition vector.

Step 20415, obtaining the text semantic cost according to the first debug cost, the second debug cost and the third debug cost.

The steps 20411-20415 are applied, the first debugging cost is obtained based on the difference between the target text semantic feature and the text semantic feature prediction result, the prediction loss of the text semantic can be represented, the second debugging cost is obtained based on the difference between different text semantic adjustment vectors, the prediction loss of the semantic adjustment/feature transformation level can be represented, and the third debugging cost can reflect the prediction loss of the text semantic from different description levels, so that the text semantic cost can be comprehensively and accurately obtained by integrating the first, second and third debugging costs.

In this embodiment of the present application, the current text semantic feature is a target text semantic feature or a text semantic feature prediction result corresponding to the authenticated description layer. Based on this, in some possible embodiments, performing semantic adjustment processing on the current text semantic feature to obtain a corresponding current text semantic adjustment vector, including: downsampling the current text semantic features to obtain first text semantic features; upsampling the first text semantic features to obtain second text semantic features; the second text semantic features and the current text semantic features have consistent vector dimensions; obtaining a target semantic difference based on the difference between the current text semantic feature and the second text semantic feature; taking the first text semantic feature as an adjusted current text semantic feature, and jumping to the step of downsampling the current text semantic feature until the first text semantic feature meets debugging expectations, so as to obtain a plurality of target semantic differences with queue priorities; and obtaining the current text semantic adjustment vector through each target semantic difference. In the embodiment of the application, through downsampling and upsampling processing on the corresponding text semantic features, accurate semantic adjustment can be realized, confusion of key details in the semantic adjustment process is avoided, determination of target semantic differences is repeatedly performed on the basis, and the current text semantic adjustment vector can be accurately determined based on the target semantic differences with queue priority.

In other possible embodiments, the streaming decision component includes at least one streaming analysis component (time sequence analysis model unit) corresponding to a description level, the streaming positive tone sample includes a streaming positive tone sample corresponding to each description level, and the streaming negative tone sample includes a streaming negative tone sample corresponding to each description level. Based on this, the loading the streaming positive and negative adjustment samples into the streaming decision component in step 2042, to obtain a first recognition perspective corresponding to the streaming positive adjustment sample and a second recognition perspective corresponding to the streaming negative adjustment sample, includes: and respectively loading the streaming positive adjustment sample and the streaming negative adjustment sample of the same description level to corresponding streaming analysis components to obtain a first recognition viewpoint and a second recognition viewpoint corresponding to each description level.

In other exemplary embodiments, the method further comprises steps 301-303.

Step 301, generating a channel attention active adjustment sample of a channel branch network corresponding to the original production control state analysis network according to the target text semantic vector, and generating a channel attention active adjustment sample of the channel branch network corresponding to the channel branch network according to the target text semantic vector and the text semantic recognition vector.

In the embodiment of the application, the channel branch network can be understood as a sub-network based on information modes (such as text mode, audio mode, image mode and the like), and on the basis of the sub-network, positive/negative modulation samples of the channel attention level can be generated.

And 302, performing joint debugging on the original production control state analysis network and the channel branch network according to the target text semantic vector, the text semantic recognition vector, the channel attention active debugging sample and the channel attention passive debugging sample to obtain updated debugging cost.

In the embodiment of the application, the channel attention active adjustment sample and the channel attention passive adjustment sample are introduced to perform joint debugging on the original production control state analysis network and the channel branch network, so that updating and debugging cost can be accurately obtained.

And 303, optimizing the network variables of the original production control state analysis network and the channel branch network according to the updating and debugging cost until the network variables meet the debugging termination requirement, and obtaining the target production control state analysis network.

It can be appreciated that, by applying steps 301-303 and performing joint debugging by introducing a channel attention (information modality) mechanism, the sensitivity of the target production control state analysis network to information of different modalities can be ensured, thereby improving the network performance of the target production control state analysis network.

Based on the foregoing, for some alternative embodiments, the method further includes step 304 and step 305.

Step 304, performing joint debugging on the original production control state analysis network, the channel branch network and the flow decision component according to the target text semantic vector, the text semantic recognition vector, the channel attention positive debugging sample, the channel attention negative debugging sample, the flow positive debugging sample and the flow negative debugging sample to obtain target debugging cost.

And step 305, optimizing the network variables of the original production control state analysis network, the channel branch network and the flow decision component according to the target debugging cost until the network variables meet the debugging termination requirement, so as to obtain the target production control state analysis network.

In the embodiment of the application, the samples of the time sequence layer and the channel attention layer can be introduced into the process of joint debugging, so that the learning capacity and the analysis capacity of the target production control state analysis network for coping with different inputs are further enriched.

In some preferred embodiments, the target text semantic vector includes a minimum of two description levels of target text semantic features and the text semantic recognition vector includes a minimum of two description levels of text semantic feature predictors. Based on this, the generating, in step 301, a channel negative attention sample corresponding to the channel branch network according to the target text semantic vector and the text semantic recognition vector includes: and generating the channel attention negative adjustment sample according to the target text semantic features of the target description level in the target text semantic vector and the text semantic feature prediction results of the residual description levels in the text semantic recognition vector.

Further, there is also provided a computer-readable storage medium having stored thereon a program which, when executed by a processor, implements the above-described method.

In the embodiments provided in the present application, it should be understood that the disclosed apparatus and method may be implemented in other manners. The apparatus and method embodiments described above are merely illustrative, for example, flow diagrams and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

In addition, the functional modules in the embodiments of the present application may be integrated together to form a single part, or each module may exist alone, or two or more modules may be integrated to form a single part.

The functions, if implemented in the form of software functional modules and sold or used as a stand-alone product, may be stored in a computer-readable storage medium. Based on such understanding, the technical solution of the present application may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a network device, or the like) to perform all or part of the steps of the methods described in the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes. It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

The foregoing description is only of the preferred embodiments of the present application and is not intended to limit the same, but rather, various modifications and variations may be made by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. The state detection method for the production control of the fireproof high-elastic fabric is characterized by being applied to an AI system, and comprises the following steps:

when the report forging discrimination tag characterizes that the to-be-analyzed finished product quality test report has forging risk, fault state detection is carried out on a fireproof high-elastic fabric production control system corresponding to the to-be-analyzed finished product quality test report;

the method for debugging the target production control state analysis network comprises the following steps:

2. The method of claim 1, wherein the obtaining a set of fire-resistant high-elastic fabric quality inspection text samples and text semantic vector samples corresponding to the set of fire-resistant high-elastic fabric quality inspection text samples comprises:

3. The method of claim 1, wherein the text semantic vector examples include text semantic vector examples corresponding to each of the fire-proof high-elastic fabric quality inspection text examples in the set of fire-proof high-elastic fabric quality inspection text examples, the text semantic vector examples, the target text semantic vector, and the text semantic recognition vector each include text semantic features of at least one descriptive layer, and the key item quality inspection text examples and the fire-proof high-elastic fabric quality inspection text examples each add a quality inspection text digital authentication signature;

4. The method of claim 1, wherein the performing joint debugging on the raw production control state analysis network and the streaming decision component according to the target text semantic vector, the text semantic recognition vector, the streaming positive debugging sample and the streaming negative debugging sample to obtain a joint debugging cost comprises:

5. The method of claim 4, wherein the target text semantic vector includes target text semantic features corresponding to a minimum of two description levels, and the text semantic recognition vector includes text semantic feature predictions corresponding to the minimum of two description levels;

obtaining the text semantic cost according to the first debugging cost, the second debugging cost and the third debugging cost;

the method comprises the steps of carrying out semantic adjustment processing on the current text semantic features to obtain corresponding current text semantic adjustment vectors, wherein the current text semantic features are target text semantic features or text semantic feature prediction results corresponding to the authenticated description level, and the method comprises the following steps: downsampling the current text semantic features to obtain first text semantic features; upsampling the first text semantic features to obtain second text semantic features; the second text semantic features and the current text semantic features have consistent vector dimensions; obtaining a target semantic difference based on the difference between the current text semantic feature and the second text semantic feature; taking the first text semantic feature as an adjusted current text semantic feature, and jumping to the step of downsampling the current text semantic feature until the first text semantic feature meets debugging expectations, so as to obtain a plurality of target semantic differences with queue priorities; and obtaining the current text semantic adjustment vector through each target semantic difference.

6. The method of claim 4, wherein the streaming decision component comprises at least one descriptive layer corresponding streaming analysis component, the streaming positive tuning sample comprises a streaming positive tuning sample corresponding to each descriptive layer, and the streaming negative tuning sample comprises a streaming negative tuning sample corresponding to each descriptive layer;

7. The method of claim 1, wherein the method further comprises:

optimizing network variables of the original production control state analysis network and the channel branch network according to the updating and debugging cost until the network variables meet the debugging termination requirement, and obtaining a target production control state analysis network;

wherein the method further comprises: according to the target text semantic vector, the text semantic recognition vector, the channel attention positive adjustment sample, the channel attention negative adjustment sample, the flow positive adjustment sample and the flow negative adjustment sample, carrying out joint debugging on the original production control state analysis network, the channel branch network and the flow decision component to obtain target debugging cost; optimizing network variables of the original production control state analysis network, the channel branch network and the flow decision component according to the target debugging cost until the network variables meet the debugging termination requirement, so as to obtain a target production control state analysis network;

The target text semantic vector comprises at least two description layers of target text semantic features, and the text semantic recognition vector comprises at least two description layers of text semantic feature prediction results; the generating the channel attention negative test sample corresponding to the channel branch network according to the target text semantic vector and the text semantic recognition vector comprises the following steps: and generating the channel attention negative adjustment sample according to the target text semantic features of the target description level in the target text semantic vector and the text semantic feature prediction results of the residual description levels in the text semantic recognition vector.

8. An AI system comprising a processor and a memory; the processor is communicatively connected to the memory, the processor being configured to read a computer program from the memory and execute the computer program to implement the method of any of claims 1-7.

9. A computer readable storage medium, characterized in that a program is stored thereon, which program, when being executed by a processor, implements the method of any of claims 1-7.