918 results about "Hidden layer" patented technology

The invention relates to the technical field of air conditioning, in particular to a laboratory heating and ventilation load prediction and self-adaptive regulation and control method. According to the method, the infrared frame and the power sampling time mark are synchronized, the sensing flow is aligned and packaged, and the thermal diffusion evolution rate is generated. Lagging characteristics are determined in combination with power jump and temperature rise moments, and a heterogeneous dynamic coupling model is established. Extracting a physical evolution parameter as a mechanism operator, injecting the mechanism operator into a hidden layer of the prediction model, reconstructing a phase space, and calculating an air load increment in a lag window. And reverse mapping is executed based on the heat exchange characteristics to generate a feedforward instruction, and when the rate exceeds a threshold value, the weight is issued and dynamically corrected, so that closed-loop correction is completed. According to the method, deep coupling of feedforward prediction compensation and feedback residual adjustment is executed, and accurate regulation and control of the air volume and cooling and heating loads of the laboratory are achieved.

The invention discloses a cross-domain privacy protection method and system for advertisement recommendation, equipment and a medium, and the method specifically comprises the steps: carrying out the encryption matching of user behavior data and anonymized equipment data, and generating an initial cross-domain joint feature vector; fusing the initial cross-domain joint feature vector and the disturbance feature vector to form a target cross-domain joint feature vector; splitting a pre-trained advertisement recommendation model into a feature coding sub-module and a reasoning sub-module, deploying the feature coding sub-module to a user adjacent edge node, and retaining the reasoning sub-module in a user local device; and based on the target cross-domain joint feature vector, performing calculation of the feature coding sub-module and calculation of the reasoning sub-module, and uploading the encrypted hidden layer feature vector to a federated learning aggregation server for global model updating. According to the method, cross-domain data utilization and user privacy protection in an advertisement recommendation process are realized, and effective feature vectors are generated for personalized advertisement recommendation while data are guaranteed not to be out of a domain.

The invention discloses a large model training method, a question and answer method, related equipment and a computer program product. The method comprises the steps of obtaining question and answer training data; the problem samples are sent to a to-be-trained large model for reasoning, and prediction output of the large model is obtained; calculating an exploration reward based on hidden layer state characteristics generated in a big model reasoning process, wherein the exploration reward is used for encouraging the big model to process the problem sample by adopting an unknown reasoning path; calculating a result reward based on the predicted output and the answer tag; and updating parameters of the large model by adopting a reinforcement learning mode according to the exploration rewards and the result rewards. Exploration rewards are additionally added in the reinforcement learning process, a large model can be encouraged to explore an unknown reasoning path, local optimum is avoided, the probability that the large model discovers a better reasoning path when facing a complex problem is improved, the performance of the large model on a complex reasoning task is improved, and the reasoning efficiency of the large model is improved. And the accuracy of the answering result of the complex reasoning question is improved.

The invention relates to a trajectory prediction model robustness enhancement method based on dynamic subspace projection decomposition. Comprising the following steps: firstly, extracting hidden layer semantic features containing historical tracks and map topology through a multi-modal feature encoder; secondly, constructing a dynamic routing mechanism based on scene self-adaption, and calculating projection weights of input features on a plurality of expert subspaces; then, executing truncation projection operation based on orthogonal decomposition, retaining core semantics located in a low-dimensional space, and filtering out adversarial disturbance located in an orthogonal complementary space; and finally, introducing a feature consistency constraint training mechanism, taking the reconstructed features of the clean sample as anchor points, and compulsively aligning the purified features of the confrontation sample with the anchor points. Compared with the prior art, the method has the advantages that the robustness of the model in white box gradient attack, black box query attack and physical semantic deception scenes is remarkably improved through feature purification of a physical level and structured consistency constraint, and the prediction reliability of the automatic driving system is ensured.

The invention relates to the technical field of motor fault diagnosis, in particular to a fault diagnosis method of a drilling machine variable frequency driving motor based on IMSOA-MCNN-BIGRU, and the method comprises the steps: collecting the current data of an experiment platform motor, and adding interference to simulate the data of a real drilling machine driving motor; a population initialization strategy of an SOA algorithm is improved, opposite-reverse learning is introduced in an iteration process, and an adaptive worst resampling mechanism after stagnation monitoring is added; the size of a convolution kernel, the size of a hidden layer, an initial learning rate and L2 regularization intensity of the MCNN-BIGRU model are optimized based on IMSOA, and optimized parameters are endowed to the MCNN-BIGRU model again so as to construct an IMSOA-MCNN-BIGRU classification model; and inputting the current data into which the interference is added into the IMSOA-MCNN-BIGRU classification model to obtain a diagnosis result. According to the method, the MCNN and the BIGRU are combined and complemented, the characterization capability and the fault judgment precision of complex non-stationary signals are improved, the MCNN-BIGRU model is optimized by using the improved sea gull optimization algorithm, and the performance of the model is improved.

The invention discloses a six-dimensional force / torque sensor decoupling system and method based on a width neural network, and the method comprises the steps: in a data collection and preprocessing unit, carrying out the null drift correction and normalization processing of a collected voltage signal through a weight type calibration platform and a data collection card, and forming a network model training data set; carrying out model training on the established six-dimensional force / torque sensor decoupling model of the width neural network, and solving a mapping matrix from a hidden layer to an output layer by adopting an importance score-based neuron pruning algorithm and a Moore-Pengos generalized inverse to obtain optimal model parameters; and finally, inputting the voltage signal subjected to data preprocessing to obtain a real-time output vector, and performing reverse normalization processing to obtain a decoupled six-dimensional force / torque vector. According to the method, precise decoupling of the six-dimensional force / torque sensor is achieved, and the inter-dimensional coupling error is remarkably reduced.

A train-track-bridge coupling response prediction method based on a sparrow optimization algorithm and a long short-term memory network comprises the steps that data such as train speed, axle load, track vibration acceleration, bridge strain and environment temperature are collected in real time through a multi-source sensor, and a multivariable time series data set is constructed after wavelet denoising and standardized preprocessing; and designing an LSTM network architecture on this basis, introducing an attention mechanism to dynamically allocate feature weights of each time step so as to enhance the ability to capture key signals in the track irregularity mutation and bridge resonance interval, and adopting a sparrow optimization algorithm to globally search an optimal combination of a hidden layer neuron number, a learning rate and a time step length in order to solve the problem of LSTM hyper-parameter optimization. Through the dynamic adaptive step length strategy balance algorithm, the early-stage global exploration and later-stage local development capabilities are balanced, the local convergence defect of a traditional grid search or genetic algorithm is avoided, the calculation efficiency can be remarkably improved, errors can be reduced, and the prediction precision can be improved.

The invention provides an industrial process fault detection method based on a space-time causal diagram autoencoder, and the method comprises the steps: 1, carrying out the data preprocessing of the space-time process data of all process variables collected in the operation process of a target industrial process for the target industrial process; step 2, establishing a causal graph space-time auto-encoder CGSTAE; 3, executing a three-step causal graph structure learning algorithm to realize training of a causal graph space-time auto-encoder CGSTAE, wherein the training comprises three steps of pre-training, causal extraction and fine tuning; and step 4, obtaining a fault detection result based on hidden layer features of the causal graph space-time auto-encoder CGSTAE and residual data output by reconstruction. According to the method, effective process monitoring and fault detection are realized by constructing two statistical magnitudes in a feature space and a residual space. Compared with other methods, the fault detection method provided by the invention can improve the reliability and interpretability of industrial process monitoring.

The invention relates to the technical field of metallurgical industry simulation and intelligent control, and discloses an intelligent metallurgical process virtual simulation method and system based on digital twinning. The method comprises the following steps: acquiring multi-modal high-dimensional data in a metallurgical process and reducing dimensions to obtain a feature vector set of a hidden layer space; performing disturbance injection simulation by using the feature vector set of the hidden layer space to obtain a multi-path state sequence with time dependence; predicting the abnormal path by using a recurrent neural network to obtain a predicted state evolution trajectory; utilizing a preset inverse mapping function to obtain multi-path state representation in the physical space; screening to obtain a risk path set; key evolution nodes are extracted from the set to be processed, and a virtual simulation scene is obtained; and performing optimization simulation on preset process adjustment parameters according to the virtual simulation scene, and determining optimized parameter configuration. The method can solve the problem that it is difficult to construct a comprehensive virtual simulation scene which truly restores the physical production rule.

The invention discloses a dynamic routing parameter efficient fine tuning method and system based on LoRA-MoE, and relates to the technical field of large model fine tuning. The method comprises the following steps: firstly, constructing a heterogeneous expert architecture-based LoRA module pool based on a multi-field data set; and secondly, coding the hidden layer features of the task through a dynamic gating network, realizing continuous differentiable expert activation, and improving the balance of expert allocation by adopting temperature attenuation and entropy regularization constraint. And then, dynamically selecting and carrying out weighted fusion on a plurality of LoRA parameter increments according to task semantics in a reasoning stage, so as to realize low-cost model adaptive updating. Finally, the module pool is continuously optimized through the confusion degree and manual evaluation feedback, low-efficiency modules are automatically eliminated, and a new LoRA module is generated to maintain task coverage. According to the method, the accuracy and generalization ability of the model in a complex scene can be remarkably improved on the premise of ensuring light weight, and rapid adaptation and dynamic optimization of a large model under a low-resource condition are realized.

The invention discloses a business risk management method and device, a computer program product and electronic equipment. The method relates to the field of artificial intelligence and big data, and comprises the following steps: obtaining business data, and encoding the business data to obtain a business vector; the business vector is input into a target model to obtain a risk prediction level, the target model comprises an input layer, a hidden layer, an attention mechanism layer and an output layer, the input layer is used for extracting a state vector from the business vector, the hidden layer is used for extracting hidden features from the state vector, the attention mechanism layer is used for calculating an attention weight, and the output layer is used for outputting the attention weight; the target feature is determined through the hidden feature and the attention weight, the attention weight is determined by the similarity between the business risk index and the hidden feature and the position code of the hidden feature, and the output layer outputs a risk prediction level based on the target feature; and determining a target risk management strategy through the risk prediction level. Through the method and the device, the problem of low business risk management efficiency in related technologies is solved.

The invention relates to the technical field of data security, and discloses a risk prediction method based on mobile terminal equipment. The method comprises the following steps: acquiring a historical behavior data set of mobile terminal equipment; a dynamic input feature set and a static interference feature set are divided according to the data set, the dynamic input feature set is derived from a user operation behavior sequence updated in real time, and the static interference feature set is generated by fusing an equipment state change record and an environment parameter change track. And establishing a dynamic risk prediction model, taking the dynamic input feature set as a model input layer, taking the static interference feature set as a model hidden layer adjustment parameter, and outputting a risk probability value by means of a bidirectional feature coupling mechanism. Segmenting and intercepting the static interference feature set by adopting a sliding time window, extracting a fluctuation period and a frequency distribution rule of the interference feature, and generating an interference feature fluctuation map. And matching the current interference characteristic state based on the map, and generating a real-time risk early warning signal in combination with the risk probability value.

The invention provides a satellite communication channel prediction system and method based on a large language model, and the system comprises a preprocessing module which is used for obtaining input data, including channel information of a time-frequency domain and a time delay-Doppler domain of satellite communication in a historical time period, and carrying out the preprocessing of the input data, and obtaining the preprocessed input data; the embedding module is used for extracting channel characteristics of a time-frequency domain and a time delay-Doppler domain from the preprocessed input data, splicing the channel characteristics of the two domains, mapping the spliced channel characteristics into an embedding space of a large language model, and then carrying out position coding to obtain a channel characteristic sequence after position coding; the large language model is used for processing the channel feature sequence after position coding and taking a result output by a hidden layer of the model as a final channel feature sequence; and the output module is used for predicting the channel information of the time-frequency domain of the satellite communication in the future preset time period according to the final channel characteristic sequence.

The invention provides a gear peeling time-varying meshing stiffness prediction method and system based on a back propagation neural network, and the method comprises the steps: considering a tooth surface peeling fault based on a gear tooth bearing contact analysis method, and constructing a helical gear pair time-varying meshing stiffness calculation model; a real irregular tooth surface peeling area is fitted by adopting a least square ellipse fitting method to obtain an ellipse appearance representation, any peeling position is completely described through six key geometric parameters, the geometric parameters of the ellipse appearance are systematically traversed and fitted, and diversified peeling appearance samples are generated. Introducing a tooth profile deviation matrix corresponding to the peeling morphology sample into a tooth surface bearing contact analysis model, and constructing a training data set; and constructing a back propagation neural network model of multiple hidden layers, and performing end-to-end training by using the training data set, so that the back propagation neural network model learns a nonlinear mapping relationship from geometric parameters to a time-varying meshing stiffness curve, thereby predicting the time-varying meshing stiffness under any peeling morphology.

The invention discloses an electric meter load real-time monitoring method based on edge calculation. The method comprises the following steps: acquiring load data in real time at an electric meter end and constructing an initial hidden layer network structure; determining an initial reference threshold value of the load data change event; detecting load data change in real time by using an event triggering mechanism and generating a triggering signal; dynamically adjusting a hidden layer node structure based on real-time load data; updating the output weight matrix by adopting a sequential learning method and predicting load data in real time; identifying a key data event according to a prediction result and executing compression packaging; and uploading the packaged key data to a remote monitoring center. According to the invention, local real-time processing of load data is realized, data transmission delay is effectively reduced, and monitoring precision and real-time performance are improved.

The invention relates to the technical field of voice recognition intelligent home control, and discloses a non-specific person voice recognition intelligent switch control method and system based on deep learning. The non-specific person voice recognition intelligent switch control method is applied to intelligent switch control equipment, and specifically comprises the following steps of S101, receiving original audio signals continuously collected in a to-be-controlled environment, and preprocessing the collected original audio signals, and then a starting point and an ending point of an effective voice segment are positioned by adopting endpoint detection based on a double-threshold method and combining the characteristic parameters of the short-time energy and the short-time zero-crossing rate. A multi-layer hidden layer structure with Dropout regularization is adopted in a neural network model, the generalization ability of the model is enhanced, a context sensing mechanism is introduced into a semantic understanding module, a composite instruction containing azimuth information can be intelligently analyzed, crossing from recognition to understanding is achieved, and the method has the advantages of being high in practicability and easy to popularize. The system is ensured to maintain a high recognition rate for voice instructions of different users under different environment conditions.

The invention discloses a model illusion detection method and device based on internal state fusion and a medium, and relates to the technical field of natural language processing. The method comprises the following steps: extracting multi-modal features in a forward propagation process of a target large language model, wherein the multi-modal features comprise a hidden layer embedding feature, an attention feature, a feedforward network activation feature and a text feature; aligning the multi-modal features to the lexical element length of the generated text through an interpolation method, and calculating the weight of the position of the lexical element corresponding to the multi-modal features; performing weighted fusion on the multi-modal features according to the weights to generate a fusion feature sequence; and constructing the fusion feature sequence into a graph structure, reasoning the graph structure by using a multi-layer attention network, and outputting a lexical-level illusion probability through a classification head. According to the method, the target model parameters do not need to be modified, and high-precision detection can be completed only through single-time forward propagation.

The invention relates to the technical field of printed circuit board signal management, and discloses an intelligent management method and system for a printed circuit board signal control process, and the method comprises the steps: obtaining the noise data of a target channel of a target management object and all adjacent channels in real time, analyzing the noise data of all channels, generating a frequency domain collaborative management domain, and carrying out the frequency domain collaborative management domain; the target management object is a server printed circuit board; the method comprises the following steps: S1, analyzing noise data to generate a frequency domain collaborative management domain, S2, mapping transient current into an N * N matrix and establishing a two-layer graph convolutional network, and S4, acquiring load data in a 100ms window, processing by using a two-way long-short-term memory network of two hidden layers, and adapting to signal transmission of a server printed circuit board, so as to avoid misjudgment caused by crosstalk. Coupling of a strong electromagnetic field is suppressed, and the signal management effect and the transmission stability are improved.

The invention discloses a power load machine learning prediction system and prediction method fusing spatial-temporal characteristics, and belongs to the technical field of power load prediction. The prediction system comprises a data acquisition module, a data preprocessing module, an AC-BiLSTM prediction module, a model evaluation module and a result output module. Multi-dimensional data are constructed into a continuous feature map as input by using a time sliding window, the advantage that a CNN can effectively extract spatial features is fully played, the ability of a BiLSTM network to extract bidirectional time sequence features of sequence data and the ability of an Attention mechanism to selectively pay attention to the hidden layer state are combined, and the hidden layer state can be effectively extracted. Therefore, the time sequence attribute of the load data is fully mined, the deep-level time correlation is obtained, and the prediction requirement of the power load data is met.

The invention relates to the technical field of space structure design and parameter optimization, in particular to a neural network model-based reticulated shell structure node parameter automatic optimization method, which comprises the following steps of: acquiring an input-output data pair of a single-layer cylindrical reticulated shell constructed by aluminum alloy plate type nodes, the input parameters comprise design parameters and initial node parameters of the single-layer cylindrical reticulated shell. According to the method, node optimization data under different parameters are obtained through interaction of a genetic algorithm, finite element software and a programming tool, an example is supplemented to construct a training data set covering a common design parameter range of a project, and then a three-layer full-connection neural network model containing two hidden layers is trained; and the model precision is ensured by matching an early stop strategy and a learning rate attenuation strategy. During application, target single-layer cylindrical reticulated shell design parameters are input, the model can automatically output optimized node parameters, repeated iteration and manual intervention of a traditional method are not needed, optimization time consumption is greatly shortened, and the multi-working-condition batch optimization requirement is efficiently met.

The invention discloses an explosion shock wave parameter prediction method and system based on a physical information neural network, and the method comprises the steps: obtaining an explosion shock wave overpressure data set of a thermobaric explosive in an open space through refined numerical simulation, and dividing the overpressure data set into a training data set and a verification data set; a specific scene is used for solving an N-S equation as a physical problem, and a mass conservation equation and momentum conservation equations in the X direction and the Y direction are involved. A neural network with no less than four hidden layers and no less than 100 neurons in each layer is constructed, the hidden layers adopt a Tanh function, an output layer uses a Softplus function to ensure non-negativity for shock wave pressure P and density rho, and speeds U and V are unconstrained. The loss function comprises a data error term and a physical information error term, the total loss function is formed by adding the data error term and the physical information error term according to weights, and specific scene boundary conditions are set. A prediction model is established based on a training result, and especially in a complex fluid scene, the method can rapidly and accurately predict an explosion shock wave overpressure field and a distribution rule.

A QR Code Verification Engine provides a multi-layered security framework for generating, validating, and authenticating QR codes while preventing tampering, fraud, and unauthorized access. The system embeds a hidden security layer within the QR code using steganographic encoding or invisible watermarking techniques, ensuring detection of any modifications. The hidden layer is encrypted using asymmetric cryptography, allowing only an authorized verification system to extract and validate it. An AI-powered tamper detection module analyzes QR codes for anomalies, while cryptographic hash verification ensures integrity. The system employs biometric authentication, push notification approvals, and contextual security measures to enhance user verification. Dynamic QR codes with expiration rules prevent replay attacks. Secure offline verification allows authentication without network connectivity. The system integrates with financial platforms, web security tools, and real-time fraud detection mechanisms, ensuring a highly secure and scalable QR code validation framework for transactions, identity verification, and access control applications.

The invention provides a loan risk assessment method and device, a storage medium and electronic equipment, and is applied to the technical field of artificial intelligence. According to the method, firstly, standardization processing is performed on loan application data, a pre-trained deep learning risk assessment model is input, and a loan risk score and hidden layer representation are obtained; and through the optimized orthogonal rotation matrix, mapping the hidden layer representation to a predefined loan risk key causal variable space, and outputting a corresponding causal variable. The variables are used for performing causal intervention verification, after an intervention result is obtained, the intervention result and a loan risk score are input into a visual interface together, a layered causal graph is generated, and a decision path of the risk score is visually displayed, so that the interpretability and transparency of loan risk assessment are improved.

The invention discloses a coating weather resistance prediction system based on a multi-task supervised neural network, and the system comprises a data obtaining module which is used for obtaining the mass ratio of a three-primary-color pigment of a target coating to an acrylic acid substrate and the coating thickness; the first neural network module is used for predicting the sunlight reflectivity of the target coating; the second neural network module is used for predicting the weather resistance of the coating; the network training module is used for training the first neural network module and the second neural network module; and the network training module is also used for completing initialization of a plurality of shared hidden layers and intermediate parameter output layers through a transfer learning mode, and completing training of the second neural network through a supervised learning mode. According to the paint weather resistance prediction system based on the multi-task supervised neural network, the limitations of single prediction target and simple model structure are solved, and an intelligent solution is provided for comprehensive evaluation and optimization of paint performance.

There is provided a volume preserving trainable artificial neural network and a system and a method for building a volume preserving trainable artificial neural network. In an aspect, an artificial neural network including: an input layer to receive input data; one or more sequentially connected hidden layers, the first hidden layer connected to the input layer, to perform operations on the input data, each hidden layer including: one or more volume-preserving rotation sublayers; one or more volume-preserving permutation sublayers; one or more volume-preserving diagonal sublayers; and an activation sublayer; and a downsizing output layer connected to the activation sublayer of the last hidden layer. In some cases, the activation sublayer includes a grouped activation function acting on a grouping of input variables to the activation sublayer.

The invention discloses a skin stretch forming residual stress rapid calculation method which comprises the following steps: randomly generating a combination of a pre-stretching amount, a coating elongation rate and a friction coefficient, submitting the combination to finite element analysis software to execute batch simulation, and generating a result file named by process parameters; traversing all the result files, extracting node numbers and residual stress values, and storing the node numbers and the residual stress values as a text format data set corresponding to the process parameters; a full-connection neural network model is constructed, an input layer receives the three process parameters of the pre-stretching amount, the coating elongation and the friction coefficient, a hidden layer comprises multiple layers of neurons and adopts a ReLU activation function, and an output layer generates residual stress values of all nodes; training the neural network model by using the data set, and adjusting the network weight through an optimizer; and inputting target process parameters to the trained neural network model, and outputting residual stress calculation results of all nodes of the skin. The technical purposes of rapidness, high efficiency and low cost are achieved.

The invention provides a federal incremental modeling method based on dynamic probability optimization in a heterogeneous edge environment. The method comprises the following steps that equipment in each factory counts class density vectors after noise adding and uploads the class density vectors to a server; the server selects a plurality of candidate sets, calculates the distribution offset coefficient of each candidate device set, and selects the minimum set to participate in federal training; equipment training is selected, and optimal parameters are uploaded; the server aggregates and distributes global model parameters; the device calculates a newly-added hidden layer output according to the global parameters and uploads the newly-added hidden layer output to the server; the server performs multi-index evaluation according to the performance of the equipment, optimizes the selected probability of the equipment through an Exp3 algorithm based on multi-index evaluation and fairness constraint, and selects the equipment to participate in training based on the probability; and when the number of hidden layer nodes exceeds the limit or the residual error satisfies the expected tolerance, finishing the training, and distributing to the equipment as a soft measurement model.

The invention discloses a photovoltaic power prediction method based on improved empirical mode decomposition and an optimized long short-term memory network, and the method comprises the steps: firstly carrying out the preprocessing of abnormal value elimination, missing value filling, normalization and the like of photovoltaic power and related meteorological data, and improving the data quality; then, an improved empirical mode decomposition (EE-ANEMD) algorithm is adopted to decompose the preprocessed power sequence into a multi-scale intrinsic mode function component and a residual term, and high-frequency noise, intermediate-frequency fluctuation and a low-frequency trend are effectively separated; global optimization is carried out on the hidden layer unit number, the initial learning rate and the maximum number of training times of the LSTM network through an improved sparrow search algorithm (ISSA), finally, the optimized LSTM is utilized to carry out training prediction on each component, and results are fused and subjected to reverse normalization to obtain a final value. Experiments show that the test set RMSE of the method is reduced compared with that of a single LSTM, the mid-term prediction precision is remarkably improved, and reliable technical support is provided for power system dispatching, new energy consumption planning and photovoltaic power station operation and maintenance.

The invention discloses a fault tree Boolean function equivalent mapping method based on an untrained neural network, and relates to the field of fault tree analysis. In order to solve the problems that in the prior art, a Boolean function mapping structure is not beneficial to parallel expansion, the calculation efficiency is limited, and the Boolean function mapping structure is difficult to efficiently realize on high-parallel platforms such as a GPU, the invention provides a method for generating topological structure data by analyzing a fault tree model; the basic events, the intermediate events and the top events are mapped into neurons of an input layer, a hidden layer and an output layer respectively, a feedforward network with fixed weight and bias is constructed, and a logic activation function is defined in nodes to realize Boolean logic propagation. The input layer receives a basic event state vector, outputs a top event result through forward propagation, and can realize large-scale Boolean function mapping on a parallel platform through batch input matrixes. The method is suitable for reliability analysis, minimum cut set simplification, top event probability calculation, parallelization fault tree solving and the like of a large-scale complex system.

A machine learning model for extracting information from web pages is prepared. The preparation includes generating respective representations of a first set of web pages, including embeddings from screenshots and bounding boxes of the web pages for multi-phase training of the model. In a first phase of training of the model, multiple loss functions associated with respective prediction tasks are optimized jointly, including a markup language element prediction task and a prediction of overlap between bounding boxes and screenshot subdivisions. In a second phase of training, using output of a hidden layer of the model (whose parameters were learned in the first phase) as input, a loss function is optimized to achieve a target web information extraction objective. The trained version of the model is stored.