Industrial Internet of Things production line safety analysis method, system, equipment and medium

Abstract

The invention discloses an industrial Internet of Things production line safety analysis method, system and device and a medium, and relates to the technical field of industrial Internet of Things, and the method comprises the steps: obtaining the operation environment time sequence data of a target production line and the corresponding fault record data; through a semantic mining model included in a target production line data analysis network, respectively performing semantic coding on the operation environment time sequence data and the fault record data to obtain corresponding operation environment features and fault features; performing cross-domain association semantic coding on the operating environment features and the fault features through the semantic analysis model to obtain corresponding environment fault association features; and performing safety situation assessment on the target production line according to the environment fault correlation features through a safety assessment model to obtain a corresponding operation safety assessment result. The method has the effect of improving the accuracy of safety analysis of the production line.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial Internet of Things, and in particular to an industrial Internet of Things production line safety analysis method, system, device and medium. BACKGROUND

[0002] The rapid development of industrial Internet of Things technology provides a large amount of device operation data for production line monitoring, mainly including continuously collected operation environment time series data (such as temperature, speed, vibration, etc.) and discretely recorded device fault data. The current production line safety monitoring method mainly relies on traditional statistical analysis, threshold-based alarm systems or shallow machine learning models.

[0003] The current technology has several significant limitations: first, the environment monitoring data and fault history data are usually processed independently, lacking an effective fusion analysis mechanism. There is a natural semantic gap between the operation environment data as a continuous time series signal and the fault record as discrete event data, and it is difficult for traditional methods to establish a deep correlation model between the two. Second, in terms of feature extraction, existing methods mostly use manually designed features or simple statistics, which cannot fully capture the nonlinear variation rules and long-range dependence relationships of device states in complex industrial environments. This leads to insufficient prediction ability for potential faults, making it difficult to achieve real early warning functions. In addition, the current safety evaluation system mostly stays at the macro analysis level of the overall production line, lacking the ability to finely evaluate the safety state of a single production device. This coarse-grained evaluation method cannot meet the demand for device-level precise maintenance of modern intelligent manufacturing systems, limiting the further improvement of operation and maintenance efficiency.

[0004] In addition to the above, the current production line data analysis method also has the problem of insufficient ability to process multi-source heterogeneous data, making it difficult to simultaneously process high-dimensional time series data and discrete event data, resulting in the inability to fully utilize multi-modal data resources in the Internet of Things environment. This limitation makes it difficult for existing systems to adapt to the precise safety performance evaluation needs in complex industrial environments.

[0005] Therefore, there is an urgent need in the field of industrial production for an innovative solution that can effectively fuse multi-source data, achieve deep feature extraction, and provide safety evaluation of different granularities, in order to overcome the shortcomings of existing technology and improve the accuracy of production line safety monitoring. SUMMARY

[0006] In order to improve the accuracy of production line safety analysis, the present application provides an industrial Internet of Things production line safety analysis method, system, device and medium.

[0007] In a first aspect, the present application provides an industrial Internet of Things production line safety analysis method, which adopts the following technical solution: The industrial internet of things production line safety analysis method is applied to an industrial internet of things system, the industrial internet of things system comprises a management platform, a sensing network platform and an object platform which are sequentially communicatively connected, the method is executed by the management platform, and comprises the following steps: obtaining running environment time sequence data of a target production line and corresponding fault record data, wherein the running environment time sequence data is running environment time sequence data of at least one production device in the target production line, the running environment time sequence data at least comprises rate data and temperature data, and the fault record data is fault record data of the at least one production device in the target production line; respectively performing semantic coding on the running environment time sequence data and the fault record data by a semantic mining model included in a target production line data analysis network, to obtain corresponding running environment features and fault features, wherein the target production line data analysis network is a pre-generated neural network model, and the target production line data analysis network further comprises a semantic analysis model and a safety evaluation model; performing cross-domain associated semantic coding on the running environment features and the fault features by the semantic analysis model, to obtain corresponding environment fault associated features, wherein the environment fault associated features are used to represent associated semantic information between a running environment and a fault of a production line; performing safety situation evaluation on the target production line according to the environment fault associated features by the safety evaluation model, to obtain a corresponding running safety evaluation result, wherein the running safety evaluation result is used to represent an overall safety state of the production line and a single safety state of the at least one production device.

[0008] By adopting the technical scheme, the running environment time sequence data of the target production line and corresponding fault record data are acquired, wherein the running environment time sequence data is the running environment time sequence data of at least one production device in the target production line, the running environment time sequence data at least includes rate data and temperature data, and the fault record data is the fault record data of at least one production device in the target production line; then the semantic mining model included in the target production line data analysis network is used to respectively perform semantic coding on the running environment time sequence data and the fault record data, to obtain corresponding running environment features and fault features; wherein the target production line data analysis network is a pre-generated neural network model, and the target production line data analysis network further includes a semantic analysis model and a safety evaluation model; then the semantic analysis model is used to perform cross-domain associated semantic coding on the running environment features and the fault features, to obtain corresponding environment fault associated features, wherein the environment fault associated features are used to represent the associated semantic information between the running environment of the production line and the fault; then the safety evaluation model is used to perform safety situation evaluation on the target production line according to the environment fault associated features, to obtain corresponding running safety evaluation results, wherein the running safety evaluation results are used to represent the overall safety state of the production line and the single safety state of at least one production device; the technical scheme of the present application realizes intelligent fusion and associated analysis of multi-source industrial data through a deep neural network, effectively solves the semantic gap problem between heterogeneous data, can accurately mine the internal correlation between the running environment parameters and the device fault, and simultaneously outputs the multi-granularity safety situation evaluation of the overall production line and the single device, thereby improving the early warning accuracy of industrial safety monitoring.

[0009] Optionally, the semantic mining model includes a vectorization sub-model, a first high-level semantic extraction sub-model and a second high-level semantic extraction sub-model; the step of respectively performing semantic coding on the running environment time sequence data and the fault record data by the semantic mining model included in the target production line data analysis network to obtain corresponding running environment features and fault features includes: respectively performing distributed representation on the running environment time sequence data and the fault record data by the vectorization sub-model to obtain corresponding environment vectors and fault vectors; performing high-level semantic extraction on the environment vectors by the first high-level semantic extraction sub-model to obtain running environment features; performing high-level semantic extraction on the fault vectors by the second high-level semantic extraction sub-model to obtain fault features.

[0010] By adopting the technical scheme, in order to obtain the operation environment feature and the fault feature, the operation environment time series data and the fault record data are respectively distributed by the vectorization sub-model to obtain corresponding environment vectors and fault vectors, then the environment vectors are subjected to high-level semantic extraction by the first high-level semantic extraction sub-model to obtain the operation environment feature, and then the fault vectors are subjected to high-level semantic extraction by the second high-level semantic extraction sub-model to obtain the fault feature.

[0011] Optionally, the step of obtaining the operation environment feature by the first high-level semantic extraction sub-model on the environment vectors comprises: The environment vectors are input into the first high-level semantic extraction sub-model, wherein the first high-level semantic extraction sub-model is internally provided with a first projection path and a second projection path, and the first projection path and the second projection path are both integrated with corresponding linear transformation units and nonlinear activation units. The environment vectors are subjected to self-attention processing to obtain environment context features. The environment context features are subjected to high-order semantic projection via the first projection path to obtain first environment projection features. The environment context features are subjected to high-order semantic projection via the second projection path to obtain second environment projection features. The first environment projection features and the second environment projection features are subjected to bidirectional interaction attention fusion calculation to synthesize the operation environment feature.

[0012] By adopting the technical scheme, in order to obtain the operation environment feature, the environment vectors are input into the first high-level semantic extraction sub-model, wherein the first high-level semantic extraction sub-model is internally provided with a first projection path and a second projection path, and the first projection path and the second projection path are both integrated with corresponding linear transformation units and nonlinear activation units, then the environment vectors are subjected to self-attention processing to obtain environment context features, then the environment context features are subjected to high-order semantic projection via the first projection path to obtain first environment projection features, then the environment context features are subjected to high-order semantic projection via the second projection path to obtain second environment projection features, and then the first environment projection features and the second environment projection features are subjected to bidirectional interaction attention fusion calculation to synthesize the operation environment feature.

[0013] Optionally, the step of obtaining the fault feature by the second high-level semantic extraction sub-model on the fault vectors comprises: inputting the fault vector into the second high-level semantic extraction sub-model, wherein the second high-level semantic extraction sub-model is internally provided with a third projection path and a fourth projection path, and the third projection path and the fourth projection path are both integrated with corresponding linear transformation units and nonlinear activation units; performing self-attention processing on the fault vector to obtain a fault context feature; performing high-order semantic projection on the fault context feature via the third projection path to obtain a first fault projection feature; performing high-order semantic projection on the fault context feature via the fourth projection path to obtain a second fault projection feature; performing bidirectional interaction attention fusion calculation on the first fault projection feature and the second fault projection feature to synthesize a fault feature.

[0014] By adopting the above technical solution, in order to obtain a fault feature, a fault vector is input into a second high-level semantic extraction sub-model, wherein the second high-level semantic extraction sub-model is internally provided with a third projection path and a fourth projection path, and the third projection path and the fourth projection path are both integrated with corresponding linear transformation units and nonlinear activation units, then self-attention processing is performed on the fault vector to obtain a fault context feature, then high-order semantic projection is performed on the fault context feature via the third projection path to obtain a first fault projection feature, then high-order semantic projection is performed on the fault context feature via the fourth projection path to obtain a second fault projection feature, and then bidirectional interaction attention fusion calculation is performed on the first fault projection feature and the second fault projection feature to synthesize a fault feature.

[0015] Optionally, the semantic analysis model includes a first semantic analysis sub-model and a second semantic analysis sub-model, and the step of performing cross-domain associated semantic coding on the running environment feature and the fault feature by the semantic analysis model to obtain corresponding environment-fault associated features includes: performing cross-domain associated semantic coding on the fault feature according to the running environment feature by the first semantic analysis sub-model to obtain a first intermediate interaction feature; performing cross-domain associated semantic coding on the fault feature according to the running environment feature by the second semantic analysis sub-model to obtain a second intermediate interaction feature; performing semantic space conversion on the running environment feature to obtain a running environment intermediate feature, wherein the first intermediate interaction feature, the second intermediate interaction feature, and the running environment intermediate feature are in the same semantic space or the same semantic dimension; splicing the first intermediate interaction feature and the second intermediate interaction feature and performing weighted summation based on an attention weight coefficient to obtain a third intermediate interaction feature; The third intermediate interaction feature and the running environment intermediate feature are added by residual error, and the added result is linearly transformed and layer normalized to obtain the environment failure correlation feature.

[0016] By adopting the technical scheme, in order to obtain the environment failure correlation feature, the first semantic analysis submodel is used to perform cross-domain associated semantic coding on the failure feature according to the running environment feature to obtain the first intermediate interaction feature, and then the second semantic analysis submodel is used to perform cross-domain associated semantic coding on the failure feature according to the running environment feature to obtain the second intermediate interaction feature, the semantic space of the running environment feature is converted to obtain the running environment intermediate feature, wherein the first intermediate interaction feature, the second intermediate interaction feature and the running environment intermediate feature are in the same semantic space or the same semantic dimension, then the first intermediate interaction feature and the second intermediate interaction feature are spliced and weighted summed based on an attention weight coefficient to obtain the third intermediate interaction feature, then the third intermediate interaction feature and the running environment intermediate feature are added by residual error, and the added result is linearly transformed and layer normalized to obtain the environment failure correlation feature.

[0017] Optionally, the first semantic analysis submodel is built-in with a first semantic space transformation matrix and a second semantic space transformation matrix, and the step of obtaining the first intermediate interaction feature by performing cross-domain associated semantic coding on the failure feature according to the running environment feature by the first semantic analysis submodel comprises: the running environment feature is transformed in semantic space by the first semantic space transformation matrix to obtain a first semantic space transformation feature, wherein the first semantic space transformation matrix is used to convert the running environment feature from a current semantic space to a target semantic space; the failure feature is transformed in semantic space by the second semantic space transformation matrix to obtain a second semantic space transformation feature, wherein the second semantic space transformation matrix is used to convert the failure feature from a current semantic space to the target semantic space; a first feature dimension correlation score between the first semantic space transformation feature and the second semantic space transformation feature is determined, and the first feature dimension correlation score is normalized to obtain a first normalized correlation score; the first semantic space transformation feature is weighted summed according to the normalized correlation score to obtain the first intermediate interaction feature.

[0018] By adopting the technical scheme, the running environment feature is subjected to semantic space transformation through a first semantic space transformation matrix to obtain a first semantic space transformation feature, wherein the first semantic space transformation matrix is used to convert the running environment feature from a current semantic space to a target semantic space, then the fault feature is subjected to semantic space transformation through a second semantic space transformation matrix to obtain a second semantic space transformation feature, wherein the second semantic space transformation matrix is used to convert the fault feature from the current semantic space to the target semantic space, then a first feature dimension correlation score between the first semantic space transformation feature and the second semantic space transformation feature is determined, and the first feature dimension correlation score is normalized to obtain a first normalized correlation score, then the first semantic space transformation feature is weighted and summed according to the normalized correlation score to obtain a first intermediate interaction feature.

[0019] Optionally, the second semantic analysis sub-model is built-in with a third semantic space transformation matrix and a fourth semantic space transformation matrix, and the step of obtaining the second intermediate interaction feature by performing cross-domain associated semantic coding on the fault feature according to the running environment feature through the second semantic analysis sub-model comprises: performing semantic space transformation on the running environment feature through the third semantic space transformation matrix to obtain a third semantic space transformation feature, wherein the third semantic space transformation matrix is used to convert the running environment feature from a current semantic space to the target semantic space; performing semantic space transformation on the fault feature through the fourth semantic space transformation matrix to obtain a fourth semantic space transformation feature, wherein the fourth semantic space transformation matrix is used to convert the fault feature from the current semantic space to the target semantic space; determining a second feature dimension correlation score between the third semantic space transformation feature and the fourth semantic space transformation feature, and normalizing the second feature dimension correlation score to obtain a second normalized correlation score; weighting and summing the third semantic space transformation feature according to the second normalized correlation score to obtain the second intermediate interaction feature.

[0020] By adopting the technical scheme, in order to obtain the second intermediate interaction feature, the running environment feature is subjected to semantic space transformation through a third semantic space transformation matrix to obtain a third semantic space transformation feature, wherein the third semantic space transformation matrix is used to convert the running environment feature from a current semantic space to a target semantic space, then the fault feature is subjected to semantic space transformation through a fourth semantic space transformation matrix to obtain a fourth semantic space transformation feature, wherein the fourth semantic space transformation matrix is used to convert the fault feature from the current semantic space to the target semantic space, then a second feature dimension correlation score between the third semantic space transformation feature and the fourth semantic space transformation feature is determined, and the second feature dimension correlation score is normalized to obtain a second normalized correlation score, then the third semantic space transformation feature is weighted and summed according to the second normalized correlation score to obtain the second intermediate interaction feature.

[0021] In a second aspect, the present application also provides an industrial Internet of Things production line safety analysis system, which adopts the following technical scheme: The industrial Internet of Things production line safety analysis system comprises a management platform, a sensing network platform and an object platform which are sequentially communicatively connected, and the management platform is configured with: a data acquisition module configured to acquire running environment time series data of a target production line and corresponding fault record data, wherein the running environment time series data is running environment time series data of at least one production device in the target production line, and at least comprises rate data and temperature data, and the fault record data is fault record data of the at least one production device in the target production line; a semantic encoding module configured to perform semantic encoding on the running environment time series data and the fault record data respectively through a semantic mining model included in a target production line data analysis network, to obtain corresponding running environment features and fault features, wherein the target production line data analysis network is a pre-generated neural network model, and the target production line data analysis network further comprises a semantic analysis model and a safety evaluation model; a semantic interaction module configured to perform cross-domain associated semantic encoding on the running environment features and the fault features through the semantic analysis model, to obtain corresponding environment fault associated features, wherein the environment fault associated features are used to represent associated semantic information between the running environment of the production line and the fault; a safety evaluation module configured to perform safety situation evaluation on the target production line according to the environment fault associated features through a safety evaluation model, to obtain a corresponding running safety evaluation result, wherein the running safety evaluation result is used to represent the overall safety state of the production line and the single safety state of the at least one production device.

[0022] In a third aspect, the present application also provides a computer device, which adopts the technical scheme as follows: A computer device comprises a memory and a processor, the memory stores a computer program capable of running on the processor, and the processor implements the method in the first aspect when executing the computer program.

[0023] In a fourth aspect, the present application also provides a computer readable storage medium, which adopts the technical scheme as follows: A computer readable storage medium stores a computer program capable of being loaded and executed by a processor to implement the method in the first aspect.

[0024] To sum up, the present application at least includes the following beneficial technical effects: obtaining running environment time sequence data of a target production line and corresponding fault record data, wherein the running environment time sequence data is running environment time sequence data of at least one production device in the target production line, the running environment time sequence data at least includes rate data and temperature data, and the fault record data is fault record data of at least one production device in the target production line; then performing semantic coding on the running environment time sequence data and the fault record data respectively through a semantic mining model included in a target production line data analysis network to obtain corresponding running environment features and fault features, wherein the target production line data analysis network is a pre-generated neural network model, and the target production line data analysis network further includes a semantic analysis model and a safety evaluation model; then performing cross-domain associated semantic coding on the running environment features and the fault features through the semantic analysis model to obtain corresponding environment fault associated features, wherein the environment fault associated features are used to represent associated semantic information between the running environment of the production line and the fault; then performing safety situation evaluation on the target production line according to the environment fault associated features through the safety evaluation model to obtain corresponding running safety evaluation results, wherein the running safety evaluation results are used to represent the overall safety state of the production line and the single safety state of at least one production device; the technical scheme of the present application realizes intelligent fusion and associated analysis of multi-source industrial data through a deep neural network, effectively solves the semantic gap problem between heterogeneous data, can accurately mine the internal correlation between the running environment parameters and the device fault, and simultaneously outputs multi-granularity safety situation evaluation of the production line as a whole and the single device, thereby improving the early warning accuracy of industrial safety monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a whole flow schematic diagram of an embodiment of the present application.

[0026] Figure 2 is a structure schematic diagram of one application scenario of a system of an embodiment of the present application.

[0027] Figure 3is a structural schematic diagram of another application scenario of a system of an embodiment of the present application.

[0028] Figure 4 is a structural block diagram of a computer device of the present application. DETAILED DESCRIPTION

[0029] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0030] Embodiments of the present application disclose an industrial Internet of Things production line safety analysis method.

[0031] Referring to Figure 1 , the industrial Internet of Things production line safety analysis method is applied to an industrial Internet of Things system, the industrial Internet of Things system includes a management platform, a sensor network platform and an object platform which are sequentially communicatively connected, the method is executed by the management platform, and includes the following steps. Step S11, obtaining running environment time series data of a target production line and corresponding fault record data.

[0032] The running environment time series data is running environment time series data of at least one production device in the target production line, and the running environment time series data at least includes speed data and temperature data. The fault record data is fault record data of at least one production device in the target production line.

[0033] It should be noted that in step S11, historical and real-time data of the target production line are collected from data sources such as Internet of Things sensors and production management systems. Specifically, two types of key information are obtained: one is running environment time series data (such as sequences of parameters such as device motor speed and bearing temperature changing over time) obtained by continuous monitoring, and the other is fault record data (such as device number, fault type, occurrence time, etc.) recorded discretely. These structured and unstructured multi-modal data jointly constitute the original input for subsequent deep analysis.

[0034] Step S12, performing semantic coding on the running environment time series data and the fault record data respectively through a semantic mining model included in a target production line data analysis network, to obtain corresponding running environment features and fault features.

[0035] The target production line data analysis network is a pre-generated neural network model, and the target production line data analysis network further includes a semantic analysis model and a safety evaluation model.

[0036] It should be noted that in step S12, the pre-trained semantic mining model (usually a deep neural network such as CNN, LSTM or Transformer) is used to abstract and reduce the dimension of the above two types of heterogeneous raw data at a high level, which can encode the original and redundant time series signal (operation data) and abstract and discrete fault events (record data) into fixed-length, dense and semantic information-rich feature vectors (i.e. operation environment features and fault features). This process converts the data from the original representation space to the semantic feature space that is more suitable for machine understanding.

[0037] Step S13, through the semantic analysis model, the operation environment features and the fault features are cross-domain associated semantic coding, and the corresponding environment fault associated features are obtained.

[0038] Among them, the environment fault associated features are used to represent the associated semantic information between the operation environment and the fault of the production line.

[0039] It should be noted that in step S13, the features from different data domains (environment domain vs. fault domain) are deeply fused and analyzed, and through a special semantic analysis model (cross attention, feature splicing, etc. Mechanism), the model will automatically learn and mine the complex and nonlinear mapping relationship between the operation environment features and the fault features, and finally synthesize a new environment fault associated feature. This feature is a unified and quantitative representation, which deeply reveals the deep semantic association of "what kind of environment state pattern will trigger or predict what kind of device fault".

[0040] Step S14, through the safety evaluation model, the safety situation of the target production line is evaluated according to the environment fault associated features, and the corresponding operation safety evaluation result is obtained.

[0041] Among them, the operation safety evaluation result is used to represent the overall safety state of the production line and the single safety state of at least one production device.

[0042] It should be noted that in step S14, based on the deeply fused associated features obtained in the previous step, a safety evaluation model (which can be a classification or regression network) is used for final decision inference. The model will comprehensively evaluate the safety status of the production system. The output result of the operation safety evaluation result contains two levels: one is the overall safety state of the production line (macro risk level), and the other is the safety state of each single production device (micro risk positioning), thereby providing the operation and maintenance personnel with a decision basis from the global to the local, hierarchical decision basis, realizing precise predictive maintenance.

[0043] In the above embodiment, the running environment time series data of the target production line and the corresponding fault record data are acquired, wherein the running environment time series data is the running environment time series data of at least one production device in the target production line, and at least includes rate data and temperature data; the fault record data is the fault record data of at least one production device in the target production line; then the semantic mining model included in the target production line data analysis network is used to respectively perform semantic coding on the running environment time series data and the fault record data, to obtain corresponding running environment features and fault features; wherein the target production line data analysis network is a pre-generated neural network model, and further includes a semantic analysis model and a safety evaluation model; then the semantic analysis model is used to perform cross-domain associated semantic coding on the running environment features and the fault features, to obtain corresponding environment fault associated features, wherein the environment fault associated features are used to represent the associated semantic information between the running environment of the production line and the fault; then the safety evaluation model is used to perform safety situation evaluation on the target production line according to the environment fault associated features, to obtain corresponding running safety evaluation results, wherein the running safety evaluation results are used to represent the overall safety state of the production line and the single safety state of at least one production device; the technical scheme of the present application realizes intelligent fusion and associated analysis of multi-source industrial data through a deep neural network, effectively solves the semantic gap problem between heterogeneous data, can accurately mine the internal correlation between the running environment parameters and the device fault, and simultaneously outputs the multi-granularity safety situation evaluation of the production line as a whole and the single device, thereby improving the early warning accuracy of industrial safety monitoring.

[0044] As a further embodiment of the method, the semantic mining model includes a vectorization sub-model, a first high-level semantic extraction sub-model and a second high-level semantic extraction sub-model; the step of respectively performing semantic coding on the running environment time series data and the fault record data through the semantic mining model included in the target production line data analysis network to obtain corresponding running environment features and fault features includes: In step S21, the vectorization sub-model is used to respectively perform distributed representation on the running environment time series data and the fault record data, to obtain corresponding environment vectors and fault vectors.

[0045] It should be noted that in step S21, the original, heterogeneous input data (numerical time series data and symbolic fault record) is converted into a standard numerical form that can be directly processed by a deep learning model; the vectorization sub-model (usually using a neural network structure such as an embedding layer, a one-dimensional convolution layer or an encoder) maps each data point or data sequence into a dense vector in a high-dimensional space (i.e. environment vector and fault vector) through distributed representation technology, thereby realizing data normalization and embedding preliminary semantic information in the vector.

[0046] Step S22, high-level semantic extraction is performed on the environment vector by a first high-level semantic extraction sub-model to obtain an operating environment feature.

[0047] It should be noted that in step S22, more abstract and representative high-level features are mined from the preliminary vectorized environment data. The first high-level semantic extraction sub-model (which can be a complex structure including a self-attention mechanism, a Transformer module or a deep convolutional network) performs a deep nonlinear transformation on the environment vector, which can capture complex temporal dependencies and internal patterns between environment parameters, filter out redundant noise, and finally output a highly refined and information-rich operating environment feature. The feature represents the deep state of the environment operation rather than the superficial raw readings.

[0048] Step S23, high-level semantic extraction is performed on the fault vector by a second high-level semantic extraction sub-model to obtain a fault feature.

[0049] It should be noted that step S23 is parallel to step S22, and the core goal is to perform high-level semantic abstraction on the fault data. The second high-level semantic extraction sub-model (which can have the same structure as the first sub-model or be specially designed to adapt to the characteristics of the fault data) is responsible for processing the fault vector. It learns the internal relationships, co-occurrence rules and deep semantics of fault events through a deep neural network, converts discrete fault records into a continuous and computable fault feature vector, and the feature can accurately express rich information such as the severity and type attributes of various faults.

[0050] In the above embodiment, in order to obtain the operating environment feature and the fault feature, the vectorization sub-model is used to perform distributed representation on the operating environment time series data and the fault record data respectively to obtain the corresponding environment vector and fault vector. Then, the first high-level semantic extraction sub-model is used to perform high-level semantic extraction on the environment vector to obtain the operating environment feature. Then, the second high-level semantic extraction sub-model is used to perform high-level semantic extraction on the fault vector to obtain the fault feature.

[0051] As a further embodiment of the method, the step of performing high-level semantic extraction on the environment vector by the first high-level semantic extraction sub-model to obtain the operating environment feature includes: Step S31, inputting the environment vector into the first high-level semantic extraction sub-model, wherein the first high-level semantic extraction sub-model has a first projection path and a second projection path, and the first projection path and the second projection path each have a corresponding linear transformation unit and a nonlinear activation unit.

[0052] Step S32, performing self-attention processing on the environment vector to obtain an environment context feature.

[0053] It should be noted that step S32 can capture the global dependency and internal structure between different elements in the environment vector. Through the self-attention processing mechanism, the model dynamically calculates the correlation weight of each time step or feature dimension in the vector with all other parts, thereby highlighting the key information and suppressing the redundant information. The output environment context feature is an enhanced feature representation that integrates global context information. Instead of being a single isolated point, it contains deep semantic information related to the elements.

[0054] Step S33, via the first projection path, performs high-level semantic projection on the environment context feature to obtain the first environment projection feature.

[0055] It should be noted that step S33 performs a deeper transformation on the context information-rich feature through the first projection path. This path projects the environment context feature from the original high-dimensional space to a new subspace through a linear transformation unit (which projects and scales the feature in a specific direction) and a nonlinear activation unit (which nonlinearly maps the projection result). The first environment projection feature is obtained, which represents the environmental semantic information interpreted from a certain specific abstract angle or perspective (determined by the parameters of the first path).

[0056] Step S34, via the second projection path, performs high-level semantic projection on the environment context feature to obtain the second environment projection feature.

[0057] It should be noted that step S33 is parallel to step S33, but processes the same environment context feature through the second projection path (which has different parameters from the first path). Due to the difference in the internal transformation matrix, it will interpret and project the feature from another complementary perspective, thereby obtaining the second environment projection feature, which provides environmental semantic information from another abstract angle that is different from the first feature.

[0058] Step S35, bidirectional interactive attention fusion calculation is performed on the first environment projection feature and the second environment projection feature to synthesize the running environment feature.

[0059] It should be noted that step S35 can combine the complementary information extracted by the two paths. Through bidirectional interactive attention fusion calculation, the first environment projection feature and the second environment projection feature are interacted, guided and weighted each other, and finally fused into a comprehensive, stable and highly condensed running environment feature. The feature integrates the advantages of the double-path and is the highest level semantic abstraction of the original environment vector.

[0060] In the above embodiment, in order to obtain the running environment feature, the environment vector is input into the first high-level semantic extraction sub-model, wherein the first high-level semantic extraction sub-model is internally provided with a first projection path and a second projection path, the first projection path and the second projection path are both integrated with corresponding linear transformation units and nonlinear activation units, then the environment vector is subjected to self-attention processing to obtain an environment context feature, then the environment context feature is subjected to high-order semantic projection through the first projection path to obtain a first environment projection feature, then the environment context feature is subjected to high-order semantic projection through the second projection path to obtain a second environment projection feature, and then bidirectional interaction attention fusion calculation is performed on the first environment projection feature and the second environment projection feature to synthesize the running environment feature.

[0061] As a further embodiment of the method, the step of extracting high-level semantics of the fault vector by the second high-level semantic extraction sub-model to obtain the fault feature comprises: Step S41, inputting the fault vector into the second high-level semantic extraction sub-model, wherein the second high-level semantic extraction sub-model is internally provided with a third projection path and a fourth projection path, and the third projection path and the fourth projection path are both integrated with corresponding linear transformation units and nonlinear activation units.

[0062] Step S42, performing self-attention processing on the fault vector to obtain a fault context feature.

[0063] Step S43, performing high-order semantic projection on the fault context feature through the third projection path to obtain a first fault projection feature.

[0064] Step S44, performing high-order semantic projection on the fault context feature through the fourth projection path to obtain a second fault projection feature.

[0065] Step S45, performing bidirectional interaction attention fusion calculation on the first fault projection feature and the second fault projection feature to synthesize the fault feature.

[0066] It should be noted that the principles of steps S41 to S45 and steps S31 to S35 are basically the same.

[0067] In the above embodiment, in order to obtain the fault feature, the fault vector is input into the second high-level semantic extraction sub-model, wherein the second high-level semantic extraction sub-model is internally provided with a third projection path and a fourth projection path, the third projection path and the fourth projection path are integrated with corresponding linear transformation units and nonlinear activation units, then the fault vector is subjected to self-attention processing to obtain a fault context feature, then the fault context feature is subjected to high-order semantic projection through the third projection path to obtain a first fault projection feature, then the fault context feature is subjected to high-order semantic projection through the fourth projection path to obtain a second fault projection feature, and then bidirectional interaction attention fusion calculation is performed on the first fault projection feature and the second fault projection feature to synthesize the fault feature.

[0068] As a further embodiment of the method, the semantic analysis model comprises a first semantic analysis sub-model and a second semantic analysis sub-model, and the step of performing cross-domain associated semantic coding on the operation environment feature and the fault feature by the semantic analysis model to obtain corresponding environment-fault associated features comprises: In step S51, the first semantic analysis sub-model is used to perform cross-domain associated semantic coding on the operation environment feature and the fault feature to obtain a first intermediate interaction feature.

[0069] It should be noted that step S51 establishes a deep association between the operation environment and the fault from a first specific perspective, and the first semantic analysis sub-model (usually a neural network based on cross-attention mechanism) receives the operation environment feature and the fault feature as input, and its core function is to take the fault feature as a query (Query) and the operation environment feature as a key and a value (Key / Value), and to mine the "environment state feature corresponding to the fault mode" by calculating the cross-attention between the two, and the output first intermediate interaction feature of the model is a quantitative representation of the environment-fault association relationship interpreted from the first semantic perspective.

[0070] In step S52, the second semantic analysis sub-model is used to perform cross-domain associated semantic coding on the operation environment feature and the fault feature to obtain a second intermediate interaction feature.

[0071] It should be noted that step S51 mines the environment-fault association relationship from another complementary perspective, and the second semantic analysis sub-model (having different parameters from the first sub-model) takes the same operation environment feature and fault feature as input, but may use different internal mapping or attention calculation methods, and it analyzes the same input from a second independently learned perspective to output a second intermediate interaction feature, which provides another interpretation of the environment-fault association relationship, and forms a complement to the first intermediate feature, thereby enhancing the representation ability and robustness of the model.

[0072] In step S53, the running environment feature is subjected to semantic space conversion to obtain a running environment intermediate feature, wherein the first intermediate interaction feature, the second intermediate interaction feature, and the running environment intermediate feature are in the same semantic space or the same semantic dimension.

[0073] It should be noted that, since the running environment feature is originally in its own semantic space, and the first intermediate interaction feature and the second intermediate interaction feature are features generated by the semantic analysis sub-model in its specific parameter space, in order to ensure the mathematical rationality of subsequent operations, the running environment feature needs to be projected into a unified target semantic space through a semantic space conversion operation to obtain the running environment intermediate feature, so as to ensure that all features to be fused (the first intermediate interaction feature, the second intermediate interaction feature, and the running environment intermediate feature) are in the same semantic space or have the same semantic dimension.

[0074] In step S54, the first intermediate interaction feature and the second intermediate interaction feature are spliced and weighted summed based on an attention weight coefficient to obtain a third intermediate interaction feature.

[0075] It should be noted that, in step S54, the first intermediate interaction feature and the second intermediate interaction feature are spliced to form a comprehensive feature representation, and then an attention weight coefficient generation mechanism (such as a small neural network or dot product attention) is used to automatically calculate and learn the importance weight of the two features in the final fusion result, and finally the spliced feature is weighted summed according to the weight to obtain the third intermediate interaction feature, which integrates the correlation information of the two perspectives and is a more comprehensive and richer representation of the environment-fault correlation.

[0076] In step S55, the third intermediate interaction feature and the running environment intermediate feature are subjected to residual addition, and the addition result is subjected to linear transformation and layer normalization to obtain an environment-fault correlation feature.

[0077] It should be noted that, in step S55, the fused cross-domain correlation feature (the third intermediate interaction feature) and the spatially unified original environment information (the running environment intermediate feature) are subjected to residual addition, which not only protects the original environment information from being lost in complex transformation, but also allows the model to learn residual changes based on the original information, effectively alleviating the gradient vanishing problem in deep networks, making the model easier to train. Subsequently, the addition result is subjected to linear transformation (for further fusion of information and adjustment of dimensions) and layer normalization (for stabilizing data distribution and accelerating training convergence), and finally a stable and information-rich environment-fault correlation feature is output.

[0078] In the above embodiment, in order to obtain the environment-failure correlation feature, the first semantic analysis sub-model is used to perform cross-domain associated semantic coding on the failure feature according to the operating environment feature to obtain a first intermediate interaction feature, and then the second semantic analysis sub-model is used to perform cross-domain associated semantic coding on the failure feature according to the operating environment feature to obtain a second intermediate interaction feature, the operating environment feature is subjected to semantic space conversion to obtain an operating environment intermediate feature, wherein the first intermediate interaction feature, the second intermediate interaction feature and the operating environment intermediate feature are in the same semantic space or the same semantic dimension, then the first intermediate interaction feature and the second intermediate interaction feature are spliced and weighted summed based on an attention weight coefficient to obtain a third intermediate interaction feature, then the third intermediate interaction feature and the operating environment intermediate feature are subjected to residual addition, and the addition result is subjected to linear transformation and layer normalization processing to obtain the environment-failure correlation feature.

[0079] As a further embodiment of the method, the first semantic analysis sub-model is built-in with a first semantic space transformation matrix and a second semantic space transformation matrix, and the step of performing cross-domain associated semantic coding on the failure feature according to the operating environment feature by the first semantic analysis sub-model to obtain the first intermediate interaction feature includes: In step S61, the operating environment feature is subjected to semantic space transformation by the first semantic space transformation matrix to obtain a first semantic space transformation feature, wherein the first semantic space transformation matrix is used to convert the operating environment feature from a current semantic space to a target semantic space.

[0080] It should be noted that, by step S61, the operating environment feature is projected from its original feature space to a unified target semantic space that is specially optimized for correlation calculation, which is realized by the first semantic space transformation matrix (a learnable linear transformation weight matrix). The matrix functions like a translator, which translates the environment feature into a universal form that can be effectively compared with the failure feature. The first semantic space transformation feature obtained after transformation will be used as Key and Value in the subsequent attention mechanism.

[0081] In step S62, the failure feature is subjected to semantic space transformation by the second semantic space transformation matrix to obtain a second semantic space transformation feature, wherein the second semantic space transformation matrix is used to convert the failure feature from a current semantic space to a target semantic space.

[0082] It should be noted that by step S62, the fault features are projected from their original space to the same target semantic space as the environment features, which is realized by another independent second semantic space transformation matrix, which is specially learned to translate the semantics of the fault features into a form matching the transformed environment features, and the second semantic space transformed features obtained after transformation will be used as Query in the subsequent attention mechanism. At this point, the two different source features are unified in the same semantic coordinate system and have comparability.

[0083] Step S63, determine the first feature dimension correlation score between the first semantic space transformed features and the second semantic space transformed features, and normalize the first feature dimension correlation score to obtain the first normalized correlation score.

[0084] It should be noted that by step S63, the correlation strength of each element in the fault features and the environment features is quantified, which is realized by calculating the similarity (usually using dot product or additive attention, etc.) between the second semantic space transformed features (Query) and the first semantic space transformed features (Key) to obtain the original first feature dimension correlation score, and then the scores are normalized (Softmax function can be used) to convert them into a probability distribution with a total sum of 1, i.e. the first normalized correlation score, which explicitly indicates the importance weight of each environment state element for the current fault query.

[0085] Step S64, according to the normalized correlation score, weighted sum of the first semantic space transformed features is obtained to obtain the first intermediate interaction features.

[0086] It should be noted that the weight obtained in the last step (the first normalized correlation score) is used to weighted sum of the first semantic space transformed features (Value), and its physical meaning is: according to the correlation strength of the fault query and the elements of the environment, the environment information is selectively focused and summarized, and the environment information highly related to the fault is strengthened, and the irrelevant information is suppressed, and finally the first intermediate interaction features are output, which is a condensed and context-related feature vector accurately representing the semantics of "which environment state information is most critical under the current fault perspective".

[0087] In the above embodiment, the running environment feature is subjected to semantic space transformation by a first semantic space transformation matrix to obtain a first semantic space transformation feature, wherein the first semantic space transformation matrix is used to convert the running environment feature from a current semantic space to a target semantic space, then the fault feature is subjected to semantic space transformation by a second semantic space transformation matrix to obtain a second semantic space transformation feature, wherein the second semantic space transformation matrix is used to convert the fault feature from the current semantic space to the target semantic space, then a first feature dimension correlation score between the first semantic space transformation feature and the second semantic space transformation feature is determined, and the first feature dimension correlation score is normalized to obtain a first normalized correlation score, then the first semantic space transformation feature is weighted and summed according to the normalized correlation score to obtain a first intermediate interaction feature.

[0088] As a further embodiment of the method, the second semantic analysis sub-model is built-in with a third semantic space transformation matrix and a fourth semantic space transformation matrix, and the step of obtaining the second intermediate interaction feature by cross-domain associated semantic coding of the fault feature according to the running environment feature through the second semantic analysis sub-model comprises: Step S71, the running environment feature is subjected to semantic space transformation by a third semantic space transformation matrix to obtain a third semantic space transformation feature, wherein the third semantic space transformation matrix is used to convert the running environment feature from a current semantic space to a target semantic space.

[0089] Step S72, the fault feature is subjected to semantic space transformation by a fourth semantic space transformation matrix to obtain a fourth semantic space transformation feature, wherein the fourth semantic space transformation matrix is used to convert the fault feature from the current semantic space to the target semantic space.

[0090] Step S73, a second feature dimension correlation score between the third semantic space transformation feature and the fourth semantic space transformation feature is determined, and the second feature dimension correlation score is normalized to obtain a second normalized correlation score.

[0091] Step S74, the third semantic space transformation feature is weighted and summed according to the second normalized correlation score to obtain the second intermediate interaction feature.

[0092] It should be noted that the principles of steps S71 to S74 and steps S61 to S64 are basically the same.

[0093] In the above embodiment, in order to obtain the second intermediate interaction feature, the running environment feature is subjected to semantic space transformation through a third semantic space transformation matrix to obtain a third semantic space transformation feature, wherein the third semantic space transformation matrix is used to convert the running environment feature from the current semantic space to the target semantic space, then the fault feature is subjected to semantic space transformation through a fourth semantic space transformation matrix to obtain a fourth semantic space transformation feature, wherein the fourth semantic space transformation matrix is used to convert the fault feature from the current semantic space to the target semantic space, then a second feature dimension correlation score between the third semantic space transformation feature and the fourth semantic space transformation feature is determined, and the second feature dimension correlation score is normalized to obtain a second normalized correlation score, then the third semantic space transformation feature is weighted and summed according to the second normalized correlation score to obtain the second intermediate interaction feature.

[0094] The application further discloses an industrial Internet of Things production line safety analysis system.

[0095] Reference Figure 2 The industrial Internet of Things production line safety analysis system comprises a management platform, a sensing network platform and an object platform which are sequentially connected in communication, and the management platform is configured with: a data acquisition module configured to acquire running environment time series data and corresponding fault record data of a target production line, wherein the running environment time series data is running environment time series data of at least one production device in the target production line, and at least comprises rate data and temperature data, and the fault record data is fault record data of at least one production device in the target production line; a semantic encoding module configured to perform semantic encoding on the running environment time series data and the fault record data respectively through a semantic mining model included in a target production line data analysis network to obtain corresponding running environment features and fault features, wherein the target production line data analysis network is a pre-generated neural network model, and the target production line data analysis network further comprises a semantic analysis model and a safety evaluation model; a semantic interaction module configured to perform cross-domain associated semantic encoding on the running environment features and the fault features through the semantic analysis model to obtain corresponding environment fault associated features, wherein the environment fault associated features are used to represent associated semantic information between the running environment and the fault of the production line; a safety evaluation module configured to perform safety situation evaluation on the target production line according to the environment fault associated features through the safety evaluation model to obtain corresponding running safety evaluation results, wherein the running safety evaluation results are used to represent the overall safety state of the production line and the single safety state of at least one production device.

[0096] The overall framework of another application scenario of the industrial Internet of Things production line safety analysis system of the application is as shown in Figure 3As shown, the user platform, the service platform, the management platform, the sensing network platform and the object platform can be sequentially interacted to form a five-platform architecture based on the industrial Internet of Things.

[0097] Specifically, in another application scenario described above, the industrial Internet of Things production line safety analysis system includes a management platform, and the management platform is configured to: acquire running environment time series data and corresponding fault record data of a target production line, wherein the running environment time series data is running environment time series data of at least one production device in the target production line, and at least includes rate data and temperature data, and the fault record data is fault record data of at least one production device in the target production line; perform semantic coding on the running environment time series data and the fault record data respectively through a semantic mining model included in a target production line data analysis network to obtain corresponding running environment features and fault features, wherein the target production line data analysis network is a pre-generated neural network model, and the target production line data analysis network further includes a semantic analysis model and a safety evaluation model; perform cross-domain associated semantic coding on the running environment features and the fault features through the semantic analysis model to obtain corresponding environment fault associated features, wherein the environment fault associated features are used to represent associated semantic information between the running environment and the fault of the production line; and perform safety situation evaluation on the target production line according to the environment fault associated features through the safety evaluation model to obtain a corresponding running safety evaluation result, wherein the running safety evaluation result is used to represent the overall safety state of the production line and the single safety state of at least one production device.

[0098] Through the interaction between the various functional platforms of the three-platform or five-platform based industrial Internet of Things production line safety analysis system, a perfect closed-loop information running logic is established to ensure the orderly running of the sensing information and the control information, and the intelligent management of the equipment is realized.

[0099] The industrial Internet of Things production line safety analysis system of the present application can implement any one of the industrial Internet of Things production line safety analysis methods, and the specific working process of the industrial Internet of Things production line safety analysis system of the present application can refer to the corresponding process in the above-mentioned industrial Internet of Things production line safety analysis methods.

[0100] The present application also discloses a computer device.

[0101] Reference Figure 4A computer device includes a memory and a processor, the memory has a computer program stored thereon, the computer program is capable of being run on the processor, and the processor implements any one of the industrial internet of things production line safety analysis methods when executing the computer program.

[0102] The embodiments of the present application further disclose a computer readable storage medium.

[0103] A computer readable storage medium stores a computer program capable of being loaded and executed by a processor to execute any one of the industrial internet of things production line safety analysis methods.

[0104] The computer readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, device or apparatus; the program code contained in the computer readable medium can be transmitted by any appropriate medium, including but not limited to wireless, wire, optical cable, RF, etc., or any appropriate combination of the above.

[0105] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application; any feature disclosed in the specification (including the abstract and the drawings) can be replaced by other equivalent or similar features, unless specifically described. That is, each feature is only an example of a series of equivalent or similar features, unless specifically described.

Claims

1. An industrial internet of things production line safety analysis method, characterized by, The method is applied to an industrial Internet of Things system, the industrial Internet of Things system comprising a management platform, a sensor network platform and an object platform which are sequentially communicatively connected, the method being executed by the management platform and comprising: obtaining running environment time series data of a target production line and corresponding fault record data, wherein the running environment time series data is running environment time series data of at least one production device in the target production line, the running environment time series data at least comprising rate data and temperature data, and the fault record data is fault record data of the at least one production device in the target production line; respectively performing semantic coding on the running environment time series data and the fault record data by means of a semantic mining model comprised in a target production line data analysis network, to obtain corresponding running environment features and fault features, wherein the target production line data analysis network is a pre-generated neural network model, and the target production line data analysis network further comprises a semantic analysis model and a security evaluation model; respectively performing cross-domain associated semantic coding on the running environment features and the fault features by means of the semantic analysis model, to obtain corresponding environment fault associated features, wherein the environment fault associated features are used to represent associated semantic information between a running environment and a fault of a production line; performing security situation evaluation on the target production line according to the environment fault associated features by means of the security evaluation model, to obtain corresponding running security evaluation results, wherein the running security evaluation results are used to represent an overall security state of the production line and a single-body security state of the at least one production device.

2. The industrial IoT production line safety analysis method of claim 1, wherein, The semantic mining model comprises a vectorization sub-model, a first high-level semantic extraction sub-model and a second high-level semantic extraction sub-model, and the step of respectively performing semantic coding on the running environment time series data and the fault record data by means of the semantic mining model comprised in the target production line data analysis network, to obtain corresponding running environment features and fault features, comprises: respectively performing distributed representation on the running environment time series data and the fault record data by means of the vectorization sub-model, to obtain corresponding environment vectors and fault vectors; performing high-level semantic extraction on the environment vectors by means of the first high-level semantic extraction sub-model, to obtain running environment features; performing high-level semantic extraction on the fault vectors by means of the second high-level semantic extraction sub-model, to obtain fault features.

3. The industrial IoT production line safety analysis method of claim 2, wherein, The step of performing high-level semantic extraction on the environment vectors by means of the first high-level semantic extraction sub-model, to obtain running environment features, comprises: inputting the environment vectors into the first high-level semantic extraction sub-model, wherein the first high-level semantic extraction sub-model is internally provided with a first projection path and a second projection path, and the first projection path and the second projection path are both integrated with corresponding linear transformation units and nonlinear activation units; performing self-attention processing on the environment vectors, to obtain environment context features; performing high-order semantic projection on the environment context features by means of the first projection path, to obtain first environment projection features; The second environment projection feature is obtained by high-order semantic projection of the environment context feature via the second projection channel; The first environment projection feature and the second environment projection feature are subjected to bidirectional interaction attention fusion calculation to synthesize the operating environment feature.

4. The industrial IoT production line safety analysis method of claim 2, wherein, The step of extracting high-level semantics of the fault vector by the second high-level semantic extraction sub-model to obtain the fault feature comprises: The fault vector is input into the second high-level semantic extraction sub-model, wherein the second high-level semantic extraction sub-model is internally provided with a third projection channel and a fourth projection channel, and the third projection channel and the fourth projection channel are both integrated with corresponding linear transformation units and nonlinear activation units; The fault vector is subjected to self-attention processing to obtain a fault context feature; The first fault projection feature is obtained by high-order semantic projection of the fault context feature via the third projection channel; The second fault projection feature is obtained by high-order semantic projection of the fault context feature via the fourth projection channel; The first fault projection feature and the second fault projection feature are subjected to bidirectional interaction attention fusion calculation to synthesize the fault feature.

5. The industrial IoT production line safety analysis method of claim 1, wherein, The semantic analysis model comprises a first semantic analysis sub-model and a second semantic analysis sub-model, and the step of performing cross-domain associated semantic coding of the operating environment feature and the fault feature by the semantic analysis model to obtain the corresponding environment-fault associated feature comprises: The first intermediate interaction feature is obtained by performing cross-domain associated semantic coding of the fault feature according to the operating environment feature by the first semantic analysis sub-model; The second intermediate interaction feature is obtained by performing cross-domain associated semantic coding of the fault feature according to the operating environment feature by the second semantic analysis sub-model; The operating environment intermediate feature is obtained by semantic space conversion of the operating environment feature, wherein the first intermediate interaction feature, the second intermediate interaction feature and the operating environment intermediate feature are in the same semantic space or the same semantic dimension; The third intermediate interaction feature is obtained by splicing the first intermediate interaction feature and the second intermediate interaction feature and performing weighted summation based on an attention weight coefficient; The environment-fault associated feature is obtained by performing residual addition of the third intermediate interaction feature and the operating environment intermediate feature and performing linear transformation and layer normalization processing on the addition result.

6. The industrial IoT production line safety analysis method of claim 5, wherein, The first semantic analysis sub-model is internally provided with a first semantic space transformation matrix and a second semantic space transformation matrix, and the step of performing cross-domain associated semantic coding of the fault feature according to the operating environment feature by the first semantic analysis sub-model to obtain the first intermediate interaction feature comprises: The first semantic space transformation feature is obtained by performing semantic space transformation of the operating environment feature by the first semantic space transformation matrix, wherein the first semantic space transformation matrix is used to convert the operating environment feature from a current semantic space to a target semantic space; The second semantic space transformation matrix is used for converting the fault feature from a current semantic space to the target semantic space. The first feature dimension correlation score between the first semantic space transformation feature and the second semantic space transformation feature is determined, and the first feature dimension correlation score is normalized to obtain a first normalized correlation score. The first intermediate interaction feature is obtained by weighted summation of the first semantic space transformation feature according to the normalized correlation score.

7. The industrial IoT production line safety analysis method of claim 6, wherein, The second semantic analysis sub-model is built-in with a third semantic space transformation matrix and a fourth semantic space transformation matrix. The third semantic space transformation matrix is used for converting the running environment feature from a current semantic space to the target semantic space. The fourth semantic space transformation matrix is used for converting the fault feature from a current semantic space to the target semantic space. The second feature dimension correlation score between the third semantic space transformation feature and the fourth semantic space transformation feature is determined, and the second feature dimension correlation score is normalized to obtain a second normalized correlation score. The second intermediate interaction feature is obtained by weighted summation of the third semantic space transformation feature according to the second normalized correlation score.

8. An industrial internet of things production line safety analysis system, characterized by, The management platform, the sensor network platform and the object platform are sequentially communicatively connected, and the management platform is configured with: a data acquisition module configured to acquire running environment time series data of a target production line and corresponding fault record data, wherein the running environment time series data is running environment time series data of at least one production device in the target production line, and at least includes rate data and temperature data, and the fault record data is fault record data of the at least one production device in the target production line; a semantic coding module configured to code the running environment time series data and the fault record data respectively by a semantic mining model included in a target production line data analysis network to obtain corresponding running environment features and fault features, wherein the target production line data analysis network is a pre-generated neural network model, and the target production line data analysis network further includes a semantic analysis model and a safety evaluation model; a semantic interaction module configured to code the running environment features and the fault features by the semantic analysis model to obtain corresponding environment fault correlation features, wherein the environment fault correlation features are used to represent the correlation semantic information between the running environment and the fault of the production line. A safety evaluation module is configured to perform safety situation evaluation on the target production line according to environment failure correlation features by a safety evaluation model, and obtain a corresponding operation safety evaluation result, wherein the operation safety evaluation result is used to represent an overall safety state of the production line and a single safety state of the at least one production device.

9. A computer device, comprising: A computer program product comprises a memory and a processor, wherein the memory stores a computer program capable of being run on the processor, and the processor implements the method according to any one of claims 1 to 7 when running the computer program.

10. A computer-readable storage medium, characterized in that, A computer program product is stored in a memory and capable of being loaded and executed by a processor to implement the method according to any one of claims 1 to 7.

Images