Industrial tablet ai software store system based on a honkong system and management method

Abstract

The application discloses an industrial tablet AI software store system based on a Hongmeng system and a management method, and belongs to the technical field of the cross of industrial internet and artificial intelligence. Through a perception and topology management subsystem, multi-dimensional working conditions, production line task roles and neighbor connection relationships of equipment are collected, time domain and frequency domain features, user behaviors, software attributes and distributed topology information enhanced by virtual nodes are fused by a multi-modal feature coding subsystem, and then a group intelligence enhanced multi-task deep recommendation network inference subsystem outputs a download probability, an adaptability score and a safety compliance score; offline training uses an improved bat algorithm fusing a wolf swarm, an ant swarm and a fruit fly search strategy to optimize in a virtual social simulation; after protocol compatibility and safety level matching are performed by an online security filtering engine, a recommended list is sorted out by combining weight values dynamically adjusted by a distributed collaborative weight optimization subsystem. The application improves the accurate perception and safety guarantee capability for industrial site software requirements.

Description

Technical Field

[0001] This invention relates to the field of industrial internet and artificial intelligence intersection technology, specifically to an industrial tablet AI software store system and management method based on HarmonyOS. Background Technology

[0002] In industrial manufacturing scenarios, industrial tablets serve as the core terminals for interaction between field operators and information systems, and their software application ecosystem is gradually evolving towards platformization and store-like models. Operators frequently need to access software stores to obtain industrial software that matches current production tasks, equipment operating conditions, and safety regulations. However, the recommendation mechanisms of existing industrial tablet software stores are generally quite simple, often employing general collaborative filtering or rule-based matching methods based on download volume. These methods fail to effectively integrate the unique multi-dimensional time-series operating condition data, production line task role information, and distributed topological relationships between devices specific to the industrial environment. This technological limitation leads to significant discrepancies between recommendation results and actual industrial needs. For example, it cannot detect the current high-load vibration state of equipment to avoid recommending resource-intensive software, nor can it automatically filter non-compliant software based on workstation safety integrity levels. Furthermore, existing methods lack production line-level group collaborative optimization mechanisms; each tablet adjusts its recommendation strategy in isolation, making it difficult to achieve optimal overall performance. Therefore, there is still room for improvement in the accuracy, adaptability, and security of software recommendations in industrial scenarios.

[0003] Therefore, how to provide an industrial tablet AI software store system and management method based on HarmonyOS has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] The purpose of this invention is to provide an industrial tablet AI software store system and management method based on the HarmonyOS system, which improves the ability to accurately perceive and ensure the security of software needs in industrial settings.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An industrial tablet AI software store system based on HarmonyOS includes: The perception and topology management subsystem is used to collect multi-dimensional time-series data of equipment operating conditions, production line status identifiers and neighbor equipment connection relationships in real time, and generate operating condition scenario matrix, task embedding vector, role embedding vector, neighbor set and communication strength. The multimodal feature encoding subsystem receives data output from the perception and topology management subsystem, as well as user behavior embedding and software attribute embedding. It encodes working condition features through a dual-channel encoding of temporal multi-scale attenuation attention and frequency domain saliency, encodes distributed topology features through a graph attention network with virtual nodes, and performs gated fusion of each modality feature to generate a multimodal fusion feature vector. The swarm intelligence-enhanced multi-task deep recommendation network inference subsystem utilizes a network structure containing a multi-head attention adaptive cross network, a dual-path deep network, and a hierarchical dynamic routing expert system to perform forward inference on the multimodal fusion feature vector, while simultaneously outputting download probability, adaptability score, and security compliance score. An improved bat algorithm training and evaluation subsystem was developed to build a training dataset offline. The system combined a homoscedastic uncertainty weighted loss function and adopted a chaotic multimorphic adaptive bat algorithm that integrates wolf pack hierarchical division of labor, ant colony pheromone and fruit fly olfactory search to optimize network hyperparameters. The fitness was evaluated in a virtual world social simulation and finally fine-tuned to obtain model parameters and hyperparameter suitability map. An online security filtering engine is used to apply industrial protocol compatibility reasoning penalties to the adaptability score in real-time inference and to correct the security compliance score by matching the SIL level. Then, the weight vector issued by the distributed collaborative weight optimization subsystem is used to sum the three scores in a weighted manner to generate a ranked recommendation list. The distributed collaborative weight optimization subsystem is used to receive local and neighborhood recommendation effect statistics online. It iteratively searches for the optimal score weight of each device by constructing the neighborhood optimal participant attractor and using an improved quantum behavior particle swarm algorithm with adaptive shrinkage-expansion coefficients. After smoothing, the score weight is sent to the online security filtering engine.

[0006] This invention also provides a management method for an industrial tablet AI software store system based on the HarmonyOS system, applied to the aforementioned industrial tablet AI software store system based on the HarmonyOS system, comprising: Step 1: Collect multi-dimensional operating condition time-series data, production line task and role information, and neighbor device connection relationships of industrial tablets, and perform normalization preprocessing to form standardized input; Step 2: Extract multi-scale time-domain features and frequency-domain salient features from the working condition time-series data and perform gated fusion. Combine user behavior embedding, software attribute embedding, and distributed topological embedding based on graph attention coding, and obtain a multimodal fusion feature vector through cross-modal gated fusion. Step 3: Input the multimodal fusion feature vector into the swarm intelligence enhanced multi-task deep recommendation network. The swarm intelligence enhanced multi-task deep recommendation network includes an adaptive cross network in which the cross order is dynamically determined by the order controller and feature interaction is achieved through multi-head attention, a dual-path deep network that fuses general path and frequency domain enhanced path gating, and a hierarchical expert system that allocates shared experts and task-specific experts based on two levels of dynamic routing gating. Finally, the download probability, adaptability score and security compliance score are output. Step 4: In the offline training phase, a training sample set containing download, adaptation, and security compliance labels is constructed. With homoscedastic uncertainty weighted loss as the optimization objective, an improved bat algorithm that combines chaotic quantum state encoding and multiple swarm intelligence strategies is used to optimize the network hyperparameters. Among them, fitness evaluation is carried out through virtual world social simulation. After obtaining the optimal hyperparameters, the model parameters are fine-tuned. Step 5: During online reasoning, the adaptability score output in Step 3 is penalized by industrial protocol compatibility knowledge graph reasoning, and the security compliance score is corrected by SIL level matching. Then, the three corrected scores are weighted and summed using dynamically updated score weights, and a recommendation list is generated by sorting the scores according to the comprehensive score. Step 6: Through a distributed collaborative mechanism, each industrial tablet periodically calculates local recommendation performance indicators, shares indicators with its neighbors, and calculates a neighborhood weighted average to improve the quantum behavior particle swarm algorithm for searching local optimal score weights. After smoothing, the score weights are used to update the score weights in Step 5.

[0007] Furthermore, in step 1, the multi-dimensional operating condition time-series data includes the CPU temperature, memory usage, root mean square value of vibration, and network packet loss rate of the industrial tablet. The neighbor device connection relationship is obtained by discovering industrial tablets of the same type through the HarmonyOS distributed soft bus and calculating the communication strength.

[0008] Furthermore, in step 2, the multi-scale temporal features are obtained by applying one-dimensional convolutions with different kernel sizes to the working condition time series data and then using attention-weighted aggregation with time decay. The significant feature in the frequency domain is obtained by independently performing a fast Fourier transform on each working condition dimension, extracting the amplitude and frequency of the frequency component with the largest amplitude, and encoding them. The distributed topology coding uses industrial tablets as nodes, neighbor relationships and communication strength as edges, and adds virtual nodes representing the overall state of the production line. The topology embedding is generated through a multi-head graph attention network.

[0009] Furthermore, in step 3, in the adaptive cross network, each cross layer interacts with the initial input and the output of the previous layer through multiple attention heads, and the resulting multi-head outputs are spliced ​​and fused with the cross terms; the sequence controller outputs the termination probability distribution of each order, and stops the cross when the cumulative termination probability exceeds the threshold; the first level of the two-level dynamic routing gating in the hierarchical expert system is sparse gating, and the second level introduces learnable prior attention to achieve fine-tuning of expert weights.

[0010] Furthermore, in step 4, the improved bat algorithm, which encodes chaotic quantum states and integrates multiple swarm intelligence strategies, specifically includes: Initial values ​​for quantum state angles are generated using Tent chaotic mapping; The population was divided into three layers based on fitness: alpha wolf, follower wolf, and scout wolf. Different layers correspond to different morphological types and search strategies. In the search space, a pheromone matrix is ​​maintained, and elite bats release pheromones and bias the quantum rotation angle. When the global optimum is continuously stagnant, Levy escape is performed by combining the historical position of the elites and the random perturbation of the fruit fly's sense of smell.

[0011] Furthermore, in step 4, the virtual world social simulation includes: A production line simulation environment with multiple virtual industrial tablets is constructed. The virtual working conditions are generated by a temporal generative adversarial network, and distributed collaboration between virtual devices is simulated. The fitness function is used to comprehensively measure the weighted result of recommendation download rate, installation success rate, security compliance rate and model complexity.

[0012] Furthermore, in step 5, the adaptability score output in step 3 is penalized using an industrial protocol compatibility knowledge graph reasoning method, specifically as follows: Two-hop queries are performed using a pre-built industrial knowledge graph to verify the compatibility between the software-driven dependency chain and the device communication protocol. If incompatible, a penalty coefficient is applied to the adaptability score. SIL level matching compares the required security integrity level of the device with the declared security integrity level of the software. When the software level is lower than the requirement, the security compliance score is set to zero.

[0013] Furthermore, in step 6, when constructing local attractors, the improved quantum behavior particle swarm algorithm replaces some dimensional components with the globally optimal position received from neighbors with probability; its contraction-expansion coefficient is adaptively adjusted according to the difference between the current globally optimal fitness and the average fitness of the particle swarm; the search results are output after exponential smoothing to avoid weight jitter.

[0014] In summary, the present invention has at least one of the following beneficial technical effects: First, by encoding and fusing multi-modal features of multi-dimensional working condition time-series data, production line task and role information, and distributed equipment topology relationships unique to industrial scenarios, the recommendation model can fully perceive the equipment operating status and production context, thereby improving the matching degree between the recommendation results and on-site needs.

[0015] Second, a multi-head attention adaptive cross-network, a dual-path deep network, and a hierarchical dynamic routing expert system are designed in the recommendation network. These systems can automatically learn high-order interaction relationships between features and dynamically allocate expert resources, thereby improving the joint prediction accuracy of download intent, adaptability, and security compliance while maintaining the model's generalization ability.

[0016] Third, during the offline training phase, a chaotic polymorphic adaptive bat algorithm that integrates wolf pack hierarchical division of labor, ant colony pheromone mechanism and fruit fly olfactory search is used to optimize the network hyperparameters. In addition, a virtual world social simulation is used to evaluate the fitness, so that the model has better global performance and robustness in the production line group collaboration scenario.

[0017] Fourth, the online reasoning process incorporates industrial protocol compatibility knowledge graph reasoning and an automatic security integrity level matching mechanism to rule-based correction of the model's output compatibility score and security compliance score. This effectively reduces the risk of incompatible software being recommended and ensures the operational reliability of software in industrial settings.

[0018] Fifth, the distributed collaborative weight optimization mechanism based on improved quantum behavior particle swarm optimization enables each industrial panel to continuously and dynamically adjust its scoring weights through neighborhood information exchange, taking into account the individual needs of each workstation and the overall efficiency of the production line, avoiding overfitting and drastic fluctuations in weights, and achieving continuous online adaptive optimization of recommendations. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0020] Figure 2 This is a schematic diagram of the convergence curve of cross-device collaborative weight adjustment based on improved quantum behavior particle swarm optimization in Example 2. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] like Figure 1 As shown, the present invention also provides a management method for an industrial tablet AI software store system based on the HarmonyOS system, including: Step 1: Collect multi-dimensional operating condition time-series data, production line task and role information, and neighbor device connection relationships of industrial tablets, and perform normalization preprocessing to form standardized input; Step 2: Extract multi-scale time-domain features and frequency-domain salient features from the working condition time-series data and perform gated fusion. Combine user behavior embedding, software attribute embedding, and distributed topological embedding based on graph attention coding, and obtain a multimodal fusion feature vector through cross-modal gated fusion. Step 3: Input the multimodal fusion feature vector into the swarm intelligence enhanced multi-task deep recommendation network. The swarm intelligence enhanced multi-task deep recommendation network includes an adaptive cross network in which the cross order is dynamically determined by the order controller and feature interaction is achieved through multi-head attention, a dual-path deep network that fuses general path and frequency domain enhanced path gating, and a hierarchical expert system that allocates shared experts and task-specific experts based on two levels of dynamic routing gating. Finally, the download probability, adaptability score and security compliance score are output. Step 4: In the offline training phase, a training sample set containing download, adaptation, and security compliance labels is constructed. With homoscedastic uncertainty weighted loss as the optimization objective, an improved bat algorithm that combines chaotic quantum state encoding and multiple swarm intelligence strategies is used to optimize the network hyperparameters. Among them, fitness evaluation is carried out through virtual world social simulation. After obtaining the optimal hyperparameters, the model parameters are fine-tuned. Step 5: During online reasoning, the adaptability score output in Step 3 is penalized by industrial protocol compatibility knowledge graph reasoning, and the security compliance score is corrected by SIL level matching. Then, the three corrected scores are weighted and summed using dynamically updated score weights, and a recommendation list is generated by sorting the scores according to the comprehensive score. Step 6: Through a distributed collaborative mechanism, each industrial tablet periodically calculates local recommendation performance indicators, shares indicators with its neighbors, and calculates a neighborhood weighted average to improve the quantum behavior particle swarm algorithm for searching local optimal score weights. After smoothing, the score weights are used to update the score weights in Step 5.

[0023] Next, the above method will be explained in detail with specific parameter settings and examples: In step 1, multi-dimensional operating condition time-series data, production line task and role information, and neighbor device connection relationships of the industrial tablet are collected. These are then pre-processed for normalization to form standardized input, specifically: Step 1 is executed by the perception and topology management subsystem. Its function is to acquire multi-dimensional environmental information about the current industrial panel and organize it into a standardized data structure that can be directly used by the subsequent multi-modal feature encoding subsystem. This step specifically includes three parallel data acquisition sub-processes, involving multi-dimensional time-series sequences of equipment operating conditions, production line status identifiers, and neighboring equipment connection relationships, and includes necessary data preprocessing operations to ensure input quality. Specifically, it includes the following sub-steps: 1. Acquisition and preprocessing of multi-dimensional time series of equipment operating conditions The device condition acquisition module uses the system resource management interface and sensor framework of the HarmonyOS operating system to achieve a fixed sampling period. Extract raw operating condition data. Sampling period. Based on a typical data refresh rate of 1 second for industrial flat panels, the total length of the time window is... The sampling point was set to 1800 points, corresponding to 30 minutes of operational history, ensuring sufficient information while avoiding excessive storage and computational overhead. At each sampling time... The module retrieves values ​​in the following four dimensions: CPU temperature, in degrees Celsius, obtained by reading / sys / class / thermal / thermal_zone0 / temp or the HarmonyOS system-level API deviceInfo.temperature.CPU.

[0024] Memory usage rate is the ratio of used physical memory to total physical memory. It is obtained through the HarmonyOS kernel memory statistics interface and the output is a dimensionless decimal between 0 and 1.

[0025] : Root mean square value of vibration, based on onboard triaxial accelerometer The calculation is performed using acceleration signals acquired within a short time window, specifically within 200 milliseconds, before and after the initial moment. Let... For the triaxial acceleration components, first remove the DC component of gravity to obtain... Composite amplitude Then calculate within the short time window The root mean square value is used as the vibration characteristic, with units of .

[0026] Network packet loss rate: The percentage of packets with no response is calculated by sending 5 ICMP request messages to the local area network gateway through the industrial tablet. The value ranges from 0 to 1.

[0027] Thus, the original four-dimensional vector is obtained at each sampling time. .

[0028] To ensure numerical stability and eliminate dimensional differences during multimodal fusion, the sequences acquired over the entire time window are preprocessed as follows: First, outlier correction is performed individually for each dimension: if the value of a sampling point deviates from the mean of its two adjacent sampling points by more than three times the local standard deviation, that point is replaced with a linear interpolation of the two adjacent points. Then, min-max normalization is used to linearly map each dimension to an interval. The historical minimum and maximum values ​​of each dimension required for normalization are obtained offline during system initialization and updated periodically. The normalized vector is denoted as... .

[0029] Arrange all sampling points within the time window in chronological order to form a working condition scenario matrix: ; in This represents the dimension of operating condition characteristics. This matrix... That is, as the output of sub-process one, it is passed to the industrial condition feature encoding module in step 2.

[0030] Input: Raw real-time data stream from the underlying sensor and system interface.

[0031] Output: Normalized working condition matrix .

[0032] 2. Obtaining production line status indicators The production line status acquisition module reads the current production line context information of the industrial tablet through a standardized Manufacturing Execution System (MES) interface or an OPCUA server. The interface returns two deterministic data points: the current production task identifier. Production line role description .in The task number is a character type, such as WO20250506-03; For predefined, finite role strings, such as assembly station visual inspection, welding robot monitoring, or conveyor belt control. Because task identifiers and role descriptions are discrete categorical variables, the module maps them to pre-trained task embedding vectors before they enter the subsequent encoding module. and role embedding vector If the production task identifier does not exist or the interface is temporarily unreachable, the default padding vector will be used, with a missing flag appended.

[0033] Input: Manufacturing Execution System task information and equipment configuration.

[0034] Output: Current production task identifier and the corresponding task embedding vector Production Line Role Description and the corresponding character embedding vector .

[0035] 3. Neighbor device discovery and connection relationship extraction The neighbor device discovery module leverages the capabilities of the HarmonyOS distributed soft bus to perform device discovery within the same local area network (LAN) or production line mesh network. The HarmonyOS distributed soft bus provides an automatic networking and service discovery mechanism based on device IDs. This module calls the `startDiscovery` interface and filters out nodes whose device type is industrial tablet. For each discovered online device, its unique identifier and communication strength are recorded. Communication strength metrics are also included. Defined as this device within a certain period of time With neighboring devices The ratio of the number of data packets transmitted via the soft bus to the average round-trip time is used to characterize the tightness of information exchange between devices. Only communication strength greater than a preset threshold is retained. The neighbors form a hop neighbor set. And construct weighted adjacency relationships.

[0036] This adjacency relationship will then be used for graph convolution operations in the distributed topology feature encoding module in step 2. Specifically, each industrial plate is treated as a graph node, and edges connect them. The edge weight is determined. For the current device The output is its list of neighbors and corresponding... Value set.

[0037] Input: Online device broadcast information and communication statistics of HarmonyOS distributed soft bus.

[0038] Output: Set of neighboring devices and the communication strength of each neighbor .

[0039] The output of the above three sub-steps Task embedding vector With character embedding vector Neighbor set With communication strength As the final output of step 1, it is passed as is to the multimodal feature encoding subsystem in step 2. The sub-processes are independent of each other, enabling fully parallel execution to meet the low-latency requirements of recommendation responses in industrial scenarios. The entire data flow in step 1 is clear, and the input and output definitions are complete, providing a solid structured perceptual data foundation for subsequent feature encoding and recommendation inference.

[0040] In step 2, multi-scale time-domain features and frequency-domain salient features are extracted from the operating condition time-series data and then gated and fused. Combined with user behavior embedding, software attribute embedding, and distributed topological embedding based on graph attention coding, a multi-modal fused feature vector is obtained through cross-modal gated fusion, specifically: Step 2 receives the working condition scenario matrix output from Step 1. Task embedding vector Role embedding vector Neighbor device collection and communication strength In conjunction with user behavior records and software attribute information already stored in the system, deep encoding is performed on five heterogeneous modalities to ultimately generate a unified multimodal fusion feature vector. This vector will serve as the sole input to the swarm intelligence-enhanced multi-task deep recommendation network in step 3. To improve the sufficiency of feature representation and scene discrimination capability, this step proposes a new architecture for industrial condition coding and distributed topology coding, and adopts a cross-modal gating mechanism in the fusion stage, forming a tightly integrated coding pipeline. Specifically, it includes the following parts: 1. Industrial operating condition characteristic coding The industrial operating condition feature encoding module receives the operating condition scenario matrix. ,in , To capture the changing patterns of operating conditions at different time scales and take into account both time and frequency domain characteristics, a dual-channel parallel encoder was designed, including a time-domain multi-scale attenuation attention channel and a frequency-domain saliency channel, which were then adaptively fused.

[0041] (1) Temporal multi-scale decaying attention channel First, a three-layer parallel multi-scale one-dimensional convolution is constructed in the temporal domain, with convolution kernel sizes of [sizes to be filled in]. , , The step size is 1, and zero padding is used to maintain the sequence length. The number of output channels in each layer is... Let the output sequence after convolution be... , The outputs at the three scales are concatenated along the channel dimension to obtain a multi-scale feature sequence. .

[0042] An improved temporal decay attention is then applied to the sequence. A learnable scale-level query vector is defined. Calculate each time step Raw attention score: ; in The sequence feature transformation matrix, To query the vector transformation matrix, For trainable attention projection vectors, for The Okay. Decay weight attenuation coefficient Set as The closer the time step, the greater the weight. Final attention weights. for: ; Time Domain Channel Output , dimension .

[0043] (2) Frequency domain saliency channel For the working condition matrix Four-dimensional signals to Perform the Fast Fourier Transform independently to obtain the amplitude spectrum of each dimension. , After smoothing the amplitude spectrum of each dimension using a moving average, the frequency component with the largest amplitude in that dimension is extracted, and its amplitude is recorded. and normalized frequency The maximum magnitude of the four dimensions. and the corresponding normalized frequency Concatenate into a vector The frequency domain features are then output through a two-layer fully connected network. The first layer is 8 to 64, ReLU is activated; the second layer is 64 to 32, ReLU is activated.

[0044] (3) Dual-channel fusion A gating mechanism is employed to fuse time-domain and frequency-domain features. The gating coefficients are calculated as follows: ; in For the sigmoid function, For the gated fusion weight matrix, This is the gated fusion bias vector. The final operating condition embedding vector. for: ; in To project time-domain features into 32 dimensions, This is the time-domain projection matrix. Output .

[0045] This encoder integrates multi-scale time-domain evolution patterns and frequency-domain periodicity characteristics, effectively identifying whether the equipment is in a high-load oscillation state or a stable idling state, providing refined operating condition semantics for recommendations.

[0046] Input: Working condition scenario matrix .

[0047] Output: Working condition embedding vector .

[0048] 2. User behavior feature extraction The user behavior feature extraction module maintains a user embedding table, which searches for the corresponding user based on the identifier of the currently logged-in operator. A 3D vector. This embedding is obtained through pre-training, which utilizes the operator's software downloads, clicks, and rating records from the past 30 days. A negative sampling Skip-gram model is used to map the behavior sequences to a continuous space. For logged-out or new operators, a default average vector is used with a learnable new user compensation term to ensure a reasonable representation during cold starts. The module outputs the user behavior embedding vector. .

[0049] Input: Current operator ID.

[0050] Output: User behavior embedding .

[0051] 3. Software attribute feature coding The software attribute feature encoding module extracts features from the metadata description text, function tags, and minimum resource requirements of each industrial software in the candidate pool. A lightweight industrial BERT model, fine-tuned offline, is used. The software description text is encoded into a 128-dimensional vector, then compressed to 32 dimensions through a linear layer. The function tags are projected to 16 dimensions after multi-hot encoding. Resource requirements, i.e., memory and storage, are normalized and concatenated to form an 8-dimensional vector. The concatenated three parts are then passed through a fully connected layer to output the software attribute embedding. The model parameters are fixed before the product goes live in the store and are not updated during the recommendation process.

[0052] Input: Software unique identifier.

[0053] Output: Software attribute embedding .

[0054] 4. Distributed topology feature coding The distributed topology feature encoding module is based on the set of neighboring devices provided in step 1. and communication strength A device graph is constructed. To introduce a social simulation mechanism, a virtual node is added in addition to the real device nodes. This represents the overall state of the production line. Virtual nodes are bidirectionally fully connected to all real device nodes, with edge weights defined by the normalized value of each device's load condition, reflecting the device's contribution to the group's state. This design draws inspiration from the core idea of ​​social simulation in swarm intelligence, where individuals adjust their behavior by perceiving the group's state.

[0055] For real nodes, the initial input features The role embedding vector of the device (From step 1, dimension 32) and working condition embedding The concatenated data is obtained through a linear transformation, maintaining a dimension of 32. Virtual node initialization features are then performed. Let it be the zero vector.

[0056] Graph encoding employs a multi-head graph attention network with edge features. For nodes... The attention operation of a single-layer graph calculates its attention to its neighbors. Attention coefficient: ; ; in To query the transformation matrix, The key transformation matrix is... , For attention parameter vectors, A learnable projection of edge weight scalars. Using... Each attention head's output is concatenated, linearly transformed, and activated by ELU. Residual connections are then superimposed, and layer normalization is performed to obtain the node representation of that layer. Stacking layers Final output As a topology embedding. For the current device node, its topology embedding is denoted as... .

[0057] The virtual nodes, embedded in the two layers, aggregate the information of the entire graph. Although the current device does not use them directly, their participation in attention calculation allows the current device to perceive its relative position and state contribution in the production line, which is in line with the design concept of social simulation.

[0058] Input: Neighbor set Communication strength Role embedding vector Working condition embedding .

[0059] Output: Topology embedding .

[0060] 5. Multimodal gating fusion The fusion module receives the above four types of embeddings All modalities have a dimension of 32. To allow different modalities to adaptively adjust their contributions according to the current scenario, a set of cross-gating coefficients is designed. First, the joint representation is calculated: ; Then, a gating vector is generated for each modality: ; in For the first The gate weight matrix for each modality, For the first The gated bias vectors of each mode. The modulated embedding is Finally, the four elements are concatenated and passed through a linear layer to output the final multimodal fused feature vector: ; To output the projection weight matrix, To output the projection bias vector, output This vector integrates the working conditions, user preferences, software attributes, and the topological role of the device in the group. It is then subjected to a gating mechanism to suppress irrelevant modal noise and is directly used as the input to the recommendation model in step 3.

[0061] enter: .

[0062] Output: Multimodal fusion feature vector .

[0063] At this point, step 2 completes all multimodal feature encoding and fusion. The entire encoding process fully leverages the unique temporal and frequency domain dual representations and virtual node-enhanced graph topology information specific to industrial scenarios. Gated fusion ensures the stability and adaptability of the recommendation model input, laying a solid feature foundation for subsequent accurate recommendations.

[0064] In step 3, the multimodal fused feature vector is input into the swarm intelligence-enhanced multi-task deep recommendation network. This network includes an adaptive crossover network where the crossover order is dynamically determined by a sequence controller and feature interaction is achieved through multi-head attention; a dual-path deep network that fuses general paths and frequency-domain enhanced paths through gating; and a hierarchical expert system that allocates shared experts and task-specific experts based on two levels of dynamic routing gating. The final outputs are download probability, suitability score, and security compliance score, specifically: This step details the forward inference network structure of the swarm intelligence-enhanced multi-task deep recommendation network. This network receives the multimodal fused feature vector output from step 2. After multi-path parallel feature learning, high-order cross-processing, and task-specific reconstruction, the download probability is finally output simultaneously. Suitability score and safety compliance score The range of values ​​for all three is [missing information]. The overall structure consists of four tightly coupled sub-networks: a multi-head attention adaptive cross-network, a dual-path deep network, a hierarchical dynamic routing expert system, and an output projection layer. The design of each sub-network fully considers the requirements of industrial scenarios for the model's nonlinear expressiveness, generalization, and stability, ensuring that the recommendations balance accuracy and security. Specifically, it includes the following components: 1. Multi-head attention adaptive cross network Multi-head attention adaptive cross-networks aim to automatically learn arbitrary-order feature cross-combinations while avoiding parameter redundancy or insufficient representation caused by fixed-order networks. The input is... Dimension .

[0065] First, a lightweight ordering controller determines the required crossover order distribution for the current sample. The input is fed into a sequence controller consisting of two fully connected layers. The first layer has 64 neurons activated by ReLU, and the second layer outputs... The cross-order weight vector is obtained by applying softmax to logits. . Indicates the network in the 1st... The probability of termination at order 1 satisfies An adaptive layer count mechanism is employed: the output of the cross-layers is calculated sequentially, and the cumulative termination probability is maintained. ,when or The process stops when the calculation stops, and any remaining uncalculated higher-order crosses are automatically ignored. In this way, different samples can dynamically select an appropriate cross depth based on their own feature complexity, avoiding overfitting for low-complexity samples and fully utilizing the higher-order expressive power for high-complexity samples.

[0066] For the The input is a cross-layer of order 1, which is Unlike traditional DCNs that use simple linear crossover, this network introduces a multi-head attention crossover mechanism. and Through respectively Each attention head interacts. For the first... One point of attention: ; in For the first Layer The query, key, and value projection matrix of each attention head. The attention output is: ; in Concatenate the outputs of the four attention points: ; The final output of the cross layer is: ; parameter This is the cross-weight matrix. For the cross bias vector, This is the attention fusion weight matrix. Introducing a low-rank decomposition structure: ,in ,rank This decomposition significantly reduces the number of parameters and enhances generalization by constraining the information flow. Furthermore, layer normalization and Dropout are applied after each layer crossover, with a dropout rate of 0.1, mitigating gradient vanishing and overfitting.

[0067] Let the actual execution order be . The final adaptive cross-network output .

[0068] 2. Dual-path deep network Deep networks employ a dual-path design, focusing on general implicit pattern extraction and industrial-specific signal enhancement respectively. Drawing on the idea of ​​division of labor and cooperation in swarm intelligence, the two paths perform their respective functions and make collaborative decisions through gating fusion.

[0069] General path: Standard residual fully connected network, number of layers Calculations for each floor: ; , For the first Layer general path weight matrix, For the first Layer-specific path bias vectors. Each layer maintains a dimension of 128 and uses residual connections. Stable training.

[0070] Industrial signal enhancement path: Design a frequency domain sensing module to transform the frequency domain saliency features from step 2. This refers to the intermediate results of the frequency domain channel from the working condition feature encoding module in step 2, which are reused here and fused with the output of the second layer of the deep network. Specifically, this involves... After linear compression to 32 dimensions, and... Element-wise product modulation is performed, and then projected back to 128 dimensions to obtain... This path allows the network to explicitly retain the frequency domain information of equipment vibration, enhancing the perception of abnormal equipment vibration. The subsequent two layers continue with residual full connections, outputting... .

[0071] The final output of the two paths is fused using gated fusion: ; ; in It is a dual-path gated weight matrix. For a two-path gated bias vector, the output is... .

[0072] 3. Hierarchical Dynamic Routing Expert System Output of cross-network and deep network output By concatenating the elements, a shared representation is obtained. .

[0073] This system adopts a hierarchical multi-task learning architecture, consisting of a bottom-level shared expert network, task-specific experts, and a two-level dynamic routing gating network. Its design draws inspiration from the task division and collaboration principles of swarm intelligence, where different experts possess unique strengths and are combined as needed through a dynamic routing mechanism. Bottom-level settings Each expert is a shared expert network, consisting of three fully connected layers with widths of 128, 64, and 32 respectively. The activation function is GELU, and the output is a 32-dimensional vector. This represents the shared network. After performing a linear transformation to 32 dimensions, the results are fed into each shared expert's database. , .

[0074] For each of the three tasks, additional equipment is provided. Each task-specific expert has a two-layer fully connected structure, with widths of 64 and 32 respectively, and also outputs 32 dimensions. Shared and task-specific experts together constitute the expert pool. The first four are shared experts, and the last six are task-specific experts.

[0075] A two-level dynamic routing gating mechanism is introduced. The first level of gating filters information between tasks, applying a gating system to each task. Generate a sparse gated vector Calculation formula: ; in For the first The first-level gating weight matrix for each task Before keeping The maximum logits are used, and the rest are set to negative infinity before softmax, which ensures that the task only activates a small number of experts, reducing parameter competition and negative transfer. The second level of gating is the fine-tuning of expert weights within the task, defining a prior attention vector based on the task's prior knowledge. , by learnable task embedding Generated through linear projection and softmax. Final expert fusion weights. , This is a learnable scalar, initialized to 0.5, used to balance the first-level sparse gating with prior knowledge. (Task) The predicted embedding is: ; .

[0076] 4. Output projection layer For each task ,Will The final prediction is obtained through a task-specific tower network. The tower network consists of two fully connected layers: the first layer is 32→16 with GELU activation; the second layer is 16→1 with sigmoid activation. Therefore: ; in For the first The first layer weight matrix of the task tower network For the first The bias vector of the first layer of the task tower network. For the first The second layer weight vector of the task tower network, For the first The second-layer bias scalar of the task tower network. For the sigmoid function, ensure the output is in The download probability was obtained for each of the three tasks. Suitability score and safety compliance score .

[0077] Summary of forward inference of the model The input is a multimodal fusion feature vector. The output is Calculation process: Multi-head attention adaptive cross network Begin with step-by-step calculations. Each stage involves interaction through 4 attention heads. and Continue until the cumulative termination probability meets the stopping condition, then output. .

[0078] Dual-path deep networks The general path and frequency domain enhancement path are executed separately, and the gated fusion is obtained. .

[0079] splicing .

[0080] Hierarchical dynamic routing expert system based on The outputs of shared computation and task-specific experts are obtained through a two-level gating weighting process. .

[0081] Output of each mission tower .

[0082] This network architecture significantly improves the accuracy and robustness of multi-target prediction in industrial scenarios through innovative designs such as multi-head attention crossover mechanism, adaptive crossover depth, low-rank crossover matrix, dual-path frequency domain enhancement, and hierarchical dynamic routing expert system. It is fundamentally different and highly innovative compared to existing technologies. Step 4 in the next section will specifically describe how to efficiently optimize the hyperparameters in the above model using swarm intelligence algorithms.

[0083] In step 4, during the offline training phase, a training sample set containing download, adaptation, and security compliance labels is constructed. Using homoscedastic uncertainty-weighted loss as the optimization objective, an improved bat algorithm employing chaotic quantum state encoding and incorporating multiple swarm intelligence strategies is used to optimize the network hyperparameters. Fitness evaluation is performed through virtual world social simulation. After obtaining the optimal hyperparameters, fine-tuning is performed to obtain the model parameters, specifically: This step details the complete training process of a swarm intelligence-enhanced multi-task deep recommendation network, covering five stages: training dataset construction, multi-task joint loss function design, hyperparameter optimization driven by an improved bat algorithm incorporating various swarm intelligence ideas, virtual world social simulation evaluation, and final model parameter fine-tuning. The overall training framework deeply integrates swarm intelligence optimization, including the pheromone mechanism of the bat algorithm and ant colony algorithm, the hierarchical division of labor in the wolf pack algorithm, and the olfactory search strategy of the fruit fly algorithm, with deep learning training to ensure optimal model performance in industrial scenarios. Specifically, it includes the following parts: 1. Training dataset construction The training data comes from the historical operation logs of HarmonyOS industrial tablets deployed on multiple industrial production lines. Each training sample is constructed at the granularity of a single software recommendation event and includes the following components: The input feature portion reuses the output structure of steps 1 and 2. Specifically, for the first... A sample of working conditions and scenarios recorded within 30 minutes prior to the recommendation time. Production line task identifier during recommendation and role embedding vector Embedded user behavior corresponding to the current operator identifier Embedded software attributes corresponding to the recommended software identifier and the set of neighboring devices built through a distributed soft bus at that time. and communication strength The above data undergoes all encoding and fusion processes in step 2 to finally obtain a multimodal fused feature vector. , as model input.

[0084] The label section contains three monitoring signals: Download tag: 1 if the operator downloads the software within 24 hours of the recommendation, otherwise 0.

[0085] : Adaptability tag, converted from the software's running status automatically collected by the system after installation. Normal operation and resource usage within the expected range is recorded as 1.0; warnings occur but the software can still run as 0.5; installation failure or crash is recorded as 0.0.

[0086] Safety compliance label: Software SIL level meets or exceeds equipment requirements, recorded as 1; otherwise, recorded as 0.

[0087] The total number of samples was drawn from historical data of multiple factories. The data, sorted chronologically, is used as the training set, with the first 80% selected. The middle 10% is used as the validation set. The last 10% was used as the test set. In addition, a lightweight quantum set was constructed by stratifying sampling from the training set according to production line and equipment type. , scale These are used for rapid evaluation during the hyperparameter optimization phase.

[0088] 2. Design of Joint Loss Function for Multiple Tasks Swarm intelligence-enhanced multi-task deep recommendation network simultaneously predicts the outputs of three tasks. A reasonable multi-task loss mechanism needs to be designed to coordinate the learning of each objective. Binary cross-entropy loss is used for download and security compliance tasks, while mean squared error loss is used for the adaptation task. ; ; ; in The batch size is used. The three tasks have different dimensions and convergence speeds; simple linear weighting can easily lead to one task dominating training. This invention employs homoscedastic uncertainty weighting, combining the loss of each task with learnable noise parameters. Correlation. The total loss function is: ; in It is a learnable scalar with an initial value of 1.0, which is jointly optimized with the network parameters during training. This term serves as a regularization mechanism, preventing noise parameters from increasing indefinitely and causing the weights of a particular task to approach zero. This design enables the model to automatically balance the training process of the three tasks.

[0089] 3. An improved bat algorithm integrating multiple swarm intelligence concepts—the chaotic polymorphic adaptive bat algorithm. Step 3's swarm intelligence-enhanced multi-task deep recommendation network contains a large number of hyperparameters: the maximum order of the multi-head attention adaptive cross-network. The actual effective order is learned by the sequence controller, but the sequence controller's own network width, Dropout rate, and number of network layers are also factors. Low-rank decomposition rank Number of attention heads Number of experts and Top-level gate sparsity Initial learning rate Batch size Weight decay coefficient These parameters constitute the hyperparameter vector to be optimized. , .

[0090] This invention proposes a chaotic polymorphic adaptive bat algorithm that integrates multiple swarm intelligence concepts for hyperparameter optimization. This algorithm makes the following fundamental improvements over the standard bat algorithm: chaotic quantum state encoding and evolution; a morphological adaptive strategy that integrates wolf pack hierarchical division of labor and ant colony pheromone mechanisms; fitness evaluation based on virtual world social simulation; and an elite memory-guided Levy escape strategy that integrates fruit fly olfactory search.

[0091] (1) Chaotic quantum state encoding and initialization Instead of direct real-value encoding as used in the standard bat algorithm, quantum state probability amplitude encoding is adopted. Each bat... The hyperparameters are determined by a pair of quantum state amplitudes. Description of the 3D hyperparameters, satisfying Actual values ​​of hyperparameters Depend on Linearly mapped to the first Dimensional search range get: ; Initialization uses Tent chaotic mapping to generate quantum state angles The Tent mapping form is: ; Take initial value for Internal random numbers, iteration This yields a chaotic sequence, which is then mapped to angles. Quantum state initialization , Chaotic initialization guarantees the traversal and uniformity of the population in the search space, which is superior to random initialization.

[0092] (2) A morphological adaptive strategy that integrates the hierarchical division of labor in wolf packs and the pheromone mechanism in ant colonies Each bat is assigned a morphology vector. The three dimensions represent exploration tendency, development tendency, and hybrid tendency, respectively, satisfying... Morphology determines the bat's search behavior pattern. Morphological vectors are uniformly distributed during initialization.

[0093] Drawing inspiration from the hierarchical division of labor in wolf pack algorithms, the bat population was sorted by fitness and divided into three tiers: the alpha wolf tier (top 20%), the follower wolf tier (middle 60%), and the exploratory wolf tier (bottom 20%). Bats in the alpha wolf tier exhibited a more active morphology, those in the exploratory wolf tier were more exploratory, and those in the follower wolf tier maintained a mixed morphology. In each generation, bats selected different evolutionary strategies based on their morphology. like A global exploration strategy is employed: quantum rotation angle. Driven by a chaotic sequence, the rotation amplitude is large, which facilitates global search.

[0094] like A local development strategy is adopted: the quantum rotation angle is guided by the current global optimum, the rotation amplitude is small, and the search is refined.

[0095] Otherwise, a hybrid strategy is adopted: the rotation angle is a weighted combination of the two above.

[0096] Drawing further inspiration from the pheromone mechanism of the ant colony algorithm, a pheromone matrix is ​​introduced into the search space. ,in Let be the number of discretized grid points for each hyperparameter, denoted as . After each iteration, elite bats release pheromones at the hyperparameter grid points they have visited; the pheromone evaporation coefficient... In the new quantum rotation angle calculation, regions with high pheromone concentrations will receive an additional angular bias: ; in This is the step size coefficient. The pheromone influence coefficient. For the first pheromone gradient in dimension For Tent chaotic mapping values, Frequency parameters: ; , Quantum state updates employ quantum rotation gates: ; The morphology vector adaptively adjusts based on fitness changes and rank classification. If a bat's fitness increases for two consecutive generations, its morphology shifts towards an exploration-oriented type; if it stagnates for two consecutive generations, its morphology shifts towards an exploratory type. Simultaneously, bats undergoing rank changes need to adjust their morphology: bats demoted from the alpha wolf rank increase their exploratory tendency, while bats upgraded from the scout wolf rank increase their exploration tendency. The update rules are as follows: ; in The change in fitness. The target shape direction vector. For morphological learning coefficients, This represents the change in level (upgrade to +1, downgrade to -1, no change to 0). Adjust the direction vector for the level. Adjust the coefficients for the grade. Perform softmax normalization after the update.

[0097] (3) Integrating elite memory of fruit fly olfactory search to guide Levy's escape strategy When the global optimal fitness is continuous When the generation is not promoted, Levi's escape is triggered. Unlike the standard bat algorithm which randomly perturbs near the global optimum, this invention combines an elite memory pool to store the positions of the top three best bats in history and constructs the escape direction.

[0098] This algorithm further borrows from the olfactory search strategy of the fruit fly algorithm, introducing random olfactory search during the escape process. In the fruit fly algorithm, individuals perceive the direction of food sources by smell and fly randomly; this algorithm incorporates this into the escape process: randomly selecting an elite from the pool. Using the difference between the current position and the elite position as the main direction, superimposed with Levy's random step size and fruit fly olfactory random perturbation: ; Lévy distribution step size Generated using the Mantegna algorithm. , This is the Lévy scaling factor. The olfactory perturbation coefficient, It is a standard normally distributed random vector. The current global optimum is identified. If a new position has better fitness, it replaces the global optimum and resets the stagnation count. This hybrid strategy utilizes historical elite information from the population to guide large-scale jumps, while olfactory perturbation in fruit flies increases the diversity of escape directions, effectively escaping local optima.

[0099] (4) Fitness assessment based on virtual world social simulation Location of each bat Corresponding hyperparameter vector Social simulations need to be conducted within a virtual world of an industrial production line to assess fitness. Virtual world simulator The design is as follows: Build a containing A production line simulation environment using virtual industrial tablets is provided, with each virtual tablet simulating equipment at different workstations. The operational sequences of the virtual tablets are generated by a pre-trained temporal generative adversarial network (TimeGAN), trained on real-world operational data, capable of synthesizing diverse operational scenarios. Virtual devices are interconnected via a simulated distributed soft bus, establishing a connection graph that mirrors the real production line topology. The currently evaluated hyperparameters are then used. A swarm intelligence-enhanced multi-task deep recommendation network is configured on all virtual devices. A simulation cycle, corresponding to 24 hours in real time, is run in the virtual world. During this cycle, virtual operators make download decisions based on their real-world behavior. The simulation ends with: Total number of recommendations; Total downloads; The number of times the program has successfully run after installation; Number of safe and compliant installations; The fitness function is designed as follows: ; in As task weight, To prevent division by zero, This is a model complexity regularization term, which is the sum of the model capacity-related dimensions in the hyperparameter vector multiplied by a penalty coefficient. A higher fitness level indicates better overall performance of the hyperparameter combination in collaborative group scenarios.

[0100] The core of social simulation lies in the fact that devices in the virtual world influence each other's recommendation weights through a simulated distributed collaborative mechanism. Their behavior is consistent with the collective intelligence collaboration in a real production line, and their collaborative mode is similar to the mechanism of indirect communication between individuals through pheromones in ant colony algorithms. This makes fitness evaluation more forward-looking and provides a holistic view of the production line than simply calculating the loss on the validation set.

[0101] 4. Overall process of hyperparameter optimization and model training Phase 1: Hyperparameter Search of the Chaotic Polymorphic Adaptive Bat Algorithm Population size Only bats, maximum number of iterations .

[0102] Each bat's quantum state encoding corresponds to a set of hyperparameters. .

[0103] Perform the following for each generation on each bat: Decoding hyperparameters in light quantum sets Train a scaled-down version of the swarm intelligence-enhanced multi-task deep recommendation network for only 10 epochs, with a batch size of 128, using the Adam optimizer and an initial learning rate of [value missing]. The learning rate value in the text.

[0104] Deploy the trained model to a virtual world simulator. The social simulation is running in the middle.

[0105] Calculate fitness .

[0106] Sort by fitness, classify wolf pack levels, update the individual optimal position, global optimal position and elite memory pool of each bat, and update the pheromone matrix.

[0107] The quantum state is updated based on morphological type and level classification, and Levy escape detection is performed by incorporating fruit fly olfactory search.

[0108] Output the optimal hyperparameter vector at the end of the iteration. .

[0109] Phase Two: Fine-tuning Training with Full Data use Configure a swarm intelligence-enhanced multi-task deep recommendation network on the complete training set. Fine-tuning was performed on the [training platform]. The optimizer was AdamW, and the initial learning rate was [initial learning rate]. Pick The optimal learning rate is determined by cosine annealing learning rate scheduling, and the minimum learning rate is achieved. Batch size Pick The optimal batch size. Weight decay coefficient. Also from The maximum number of training epochs is 200, the early stopping patience value is 20, and the monitoring metric is the overall loss on the validation set. After each round of training, on the validation set... The model parameters with the lowest validation loss are then evaluated and saved. .

[0110] After training, on the test set The final performance metrics are calculated to confirm that the model meets the deployment requirements. The final output includes: parameters of the trained swarm intelligence-enhanced multi-task deep recommendation network model. Optimal hyperparameter vector And hyperparameter suitability maps generated in virtual world social simulations.

[0111] In summary, the input and output of step 4 are: enter: Training dataset Validation set Test set and lightweight evaluation subset ; Virtual World Simulator Includes a TimeGAN condition generator, simulated equipment topology, and simulated operator behavior distribution; Upper and lower bounds of each dimension of the hyperparameter search space ; Definition of a swarm intelligence-enhanced multi-task deep recommendation network structure; Output: Parameters of the trained swarm intelligence-enhanced multi-task deep recommendation network model ; Optimal hyperparameter vector ; The hyperparameter suitability map displays the overall fitness distribution of different hyperparameter configurations across various production line scenarios in the virtual world, guiding the rapid initialization of models for subsequent new scenarios. The above training process organically combines the pheromone mechanism of the bat algorithm and the ant colony algorithm, the hierarchical division of labor of the wolf pack algorithm, the olfactory search of the fruit fly algorithm, virtual world social simulation, and deep learning training, ensuring that the model achieves global optimal performance in the group collaboration environment of the industrial production line, fully demonstrating the high level of creativity of this invention at the training method level.

[0112] In step 5, during online inference, the adaptability score output in step 3 is penalized using an industrial protocol compatibility knowledge graph, and the security compliance score is corrected using SIL level matching. Then, the three corrected scores are weighted and summed using dynamically updated scoring weights, and a recommendation list is generated by sorting the scores according to the overall score. Specifically: This step describes the complete online reasoning process of an industrial tablet in a real-world operating environment, from sensor data collection to the generation of the final recommendation list and the completion of safety filtering, when an operator triggers a software recommendation request or the system actively pushes recommendations. This step relies on the parameters of the swarm intelligence-enhanced multi-task deep recommendation network model trained in step 4. and optimal hyperparameters The process reuses the data acquisition logic from step 1 and the feature encoding path from step 2, but uses frozen parameters for forward propagation in inference mode. The virtual world suitability map constructed in step 4 serves as a reference in this cold start scenario. The overall process comprises five stages: request triggering and data acquisition, feature encoding and model inference, protocol compatibility inference, SIL security level matching, and comprehensive scoring and ranking output. The output of each stage serves as the direct input to the next stage, forming a tight data flow loop. These will be described in detail below: 1. Request Triggering and Real-time Data Acquisition When an operator opens the AI ​​software store app or clicks the smart recommendation button on the industrial tablet, the system generates a recommendation request and records the moment of the request. The three acquisition sub-processes described in step 1 are immediately triggered in parallel: (1) Equipment operating condition data collection Obtain from HarmonyOS system interface The multi-dimensional time-series sequence of operating conditions within the 30 minutes prior to the current time. Directly read the most recent 1800 operating condition records that have been continuously cached in local memory since step 1. If the cache is less than 30 minutes due to reasons such as restarts, fill the existing data with a zero vector at the front end. Sample points are selected, and the filling flag is marked as 1. The same outlier correction and min-max normalization as in step 1 are performed on the cached data, with the normalization parameters using the historical extreme values ​​updated periodically in step 1, to obtain the working condition scenario matrix. .

[0113] (2) Production line status data collection Read the current production task identifier via OPCUA or Manufacturing Execution System interface. and production line roles And map them to their respective pre-trained embedding vectors. and If the interface is unreachable, use the previously successfully obtained embedding vector and set the missing flag to 1.

[0114] (3) Neighbor device discovery Call the HarmonyOS distributed soft bus listOnlineDevices interface to obtain a list of currently online industrial tablet devices of the same type, and filter them by communication strength. The neighbors form a neighborhood set. and the strength value of each edge. If there are currently no online neighbors, then It is an empty set.

[0115] This step outputs three parts: a normalized chemical condition scenario matrix. Task embedding vector With character embedding vector Neighbor set and corresponding communication strength This output is passed to the next stage as is.

[0116] 2. Multimodal feature encoding and model forward inference This step strictly follows the feature encoding architecture defined in step 2 and the forward structure of the swarm intelligence-enhanced multi-task deep recommendation network defined in step 3, but all model parameters are in the inference frozen state and no gradient calculation is performed.

[0117] (1) Operating condition characteristic coding Work scenario matrix The data is fed into the working condition feature encoding module. The temporal multi-scale decaying attention channel extracts multi-scale sequence features through three sets of one-dimensional convolutions. Calculate the decaying attention weighted sum Frequency domain saliency channels Perform FFT on each of the four dimensions, extracting the amplitude and normalized frequency of the frequency component with the largest amplitude in each dimension, and then obtain the result through a fully connected mapping. Dual-channel gated fusion output embedded operating conditions .

[0118] (2) User behavior feature acquisition Query the user embedding table based on the currently logged-in operator ID to obtain the user behavior embedding. For new operators, use the global average embedding plus a learnable compensation vector.

[0119] (3) Software attribute feature extraction The recommended candidate pool is maintained by the store's backend and contains all registered industrial software, typically numbering in the hundreds. For each software in the candidate pool, the software attribute feature encoding module reads its metadata, and after industrial BERT projection and resource requirement encoding, obtains the software attribute embedding. To reduce redundant calculations, the backend updates and caches the embedding vectors of all software offline once a day, allowing for direct lookup of the table by software ID during online inference.

[0120] (4) Distributed topology feature encoding With the current device's neighbor set and communication strength Using this as input, and combining the operational status embedding and role embedding of each device's cache, a structure containing virtual nodes is constructed. The graph is processed using a two-layer multi-head graph attention network to obtain the topological embedding of the current device. If the neighbor set is empty, the virtual node still exists, and the attention mechanism degenerates into self-attention plus residual, ensuring that the feature dimension remains unchanged.

[0121] (5) Multimodal gating fusion embed the above four The data is fed into a multimodal gating fusion module to obtain a multimodal fusion feature vector. For each software in the candidate pool, its The differences mean that separate generation is required for each candidate software. To improve efficiency, Calculate only once for the same request, and with each After pairing, the data are sent in batches to the fusion module, where matrix operations are used to accelerate the process.

[0122] (6) Forward computation of swarm intelligence-enhanced multi-task deep recommendation network For each candidate software take it as A swarm intelligence-enhanced multi-task deep recommendation network with frozen input. A multi-head attention adaptive crossover network outputs according to the order controller. Determine the cross depth Calculate cross representation The dual-path deep network parallel computation general path and frequency domain enhancement path are obtained through gated fusion. After concatenation, the hierarchical dynamic routing expert system outputs the expert fusion vectors for each task. The mission tower network outputs three predicted values: ; ; ; Each candidate software received a three-dimensional prediction vector. .

[0123] The output of this step is a list of three-dimensional prediction scores for all candidate software.

[0124] 3. Industrial Protocol Compatibility Knowledge Graph Reasoning The model outputs a fitness score. Building upon this foundation, the system further verifies software protocol compatibility through knowledge graph reasoning. The system pre-constructs an industrial domain knowledge graph, with entity types including Software, Driver, Protocol, and Hardware, and relation types including DEPENDS_ON (software depends on a driver), IMPLEMENTS (driver implements a protocol), and COMPATIBLE_WITH (protocol is compatible with hardware). This knowledge graph is stored in the graph database Neo4j.

[0125] For each candidate software, a Cypher two-hop inference query is performed, starting from the software, passing through the driver and protocol to reach the hardware, to verify whether there is a path compatible with the current device's communication protocol. If the number of compatible paths is 0, it means that the software's driver dependency chain has no intersection with the current device's communication protocol, posing a risk of incompatibility. At this point, [further steps are needed]. Imposing punishment: ; Punishment factor If at least one compatible path exists, then This corrected score replaces the original fit score in subsequent comprehensive calculations.

[0126] 4. Automatic matching of SIL security level The security and compatibility matching layer reads the required security integrity level from the device configuration of the industrial flat panel. The value is taken from the set {1, 2, 3, 4}, with higher values ​​indicating more stringent safety requirements. Each industrial software declares its design reachability when uploaded to the store. The level, also takes values ​​from {1,2,3,4}.

[0127] For each candidate software, a rigorous comparison is performed: if This indicates that the software's security level does not meet the device's minimum requirements. Therefore, the software's security compliance prediction value will be forcibly set to zero. Otherwise, retain the original predicted value. .

[0128] 5. Overall score and ranking output The three-dimensional score after the above security filtering The scores need to be merged into a single ranking composite score. The merging weights are derived from the optimal weight vector maintained online by the collaborative weight adjustment module described in step 6. The initial value is set by offline experience. Then, it is dynamically updated through production line swarm intelligence. The overall score is calculated as follows: ; in Index candidate software. Sort the candidate software by... Sort by highest to lowest, and extract the first few. A list of recommended items. The recommended list excludes software already installed on the current device to avoid duplicate recommendations. If the number of software programs is less than 20, no hard stuffing will be done.

[0129] (1) Special treatment for cold start For newly deployed industrial tablets, the lack of historical behavioral data leads to embedded user behavior. A global average vector is used. At this point, the system will refer to the hyperparameter suitability map generated in step 4, and retrieve the most relevant set of pre-trained model parameters from the suitability map based on the production line type, role, and mainstream operating conditions of the current equipment. The system will then select the model and weights with the highest matching degree with the current feature encoding for initialization, thereby accelerating the improvement of recommendation quality during the cold start phase.

[0130] (2) Reasoning output Finally, the recommended list The data is displayed on the user interface, with each software entry accompanied by its overall score and a summary of compatibility and security tags, for the operator to choose from. Simultaneously, the system will display the requested... The final recommendation list and the user's subsequent download behavior are all recorded in the behavior log, which will serve as the data source for the model iteration training in the future step 4.

[0131] Input: The operator triggers a recommendation request signal.

[0132] Output: Recommended software list Each entry includes a software identifier, overall score, and a summary of adaptability and security status.

[0133] The entire online inference chain, from data acquisition to security filtering, has strictly corresponding inputs and outputs at each stage. The calculation process is clear and reproducible, and model inference and security rules are decoupled and complementary, ensuring the efficiency and reliability of software recommendations in industrial settings.

[0134] In step 6, through a distributed collaborative mechanism, each industrial tablet periodically calculates its local recommendation performance index, shares the index with its neighbors, and calculates a neighborhood-weighted average to improve the quantum behavior particle swarm optimization algorithm for searching for local optimal rating weights. After smoothing, this weight is used to update the rating weights in step 5. Specifically: This step constructs an online weight optimization module, which is deployed in a distributed cluster of multiple HarmonyOS industrial tablets within the production line. It continuously adjusts the weight vector in the comprehensive scoring formula of step 5 using an improved quantum behavior particle swarm optimization algorithm. This optimizes the weighted combination of overall download success rate, adaptation success rate, and security compliance rate across the production line. The weighting process does not rely on a central server; each tablet exchanges some information via the HarmonyOS distributed soft bus, forming a decentralized, collaborative, and intelligent system that balances the individual needs of each workstation with overall production line efficiency. Specifically, this includes: 1. Optimization Problem Modeling and Particle Representation Define each industrial flat panel The problem of local weight optimization. Weight vector. Satisfying constraints ,and The goal is to minimize a global loss function, which is the aggregated recommendation results of all treadmills within the production line.

[0135] Each plate maintains a particle swarm to search for its locally optimal weights. Each particle exist The position of a time represents a set of weights. Speed ​​is denoted as All tablets share the same structure and update logic of the improved quantum behavior particle swarm optimization algorithm, but the particle swarm and fitness evaluations are based on local data and neighborhood data obtained through cooperative exchange.

[0136] 2. Fitness function design, weighted comprehensive utility at the production line level Because recommendation decisions across devices influence each other, factors such as the performance feedback of software on multiple devices and the collaborative needs of upstream and downstream production lines can lead to local overfitting of weights in single-device optimization. Therefore, the fitness function design needs to integrate local feedback and utility metrics shared by neighboring devices.

[0137] Every minutes, tablet With the neighbor set maintained by the distributed topology module in step 1 Swap the following metrics: Past Within the window, on this device: Total number of recommended displays ; Number of user downloads ; Number of times the software was successfully downloaded, installed, and ran stably. ; Number of times software is installed to meet SIL compliance ; The three local sub-indicators are calculated as follows, using exponential smoothing to fuse historical values: ; ; ; Smoothing factor , To avoid division by zero. Initial value Take 0.5.

[0138] To gain a global view of the production line, a neighborhood average metric is introduced. For each metric... Calculate the neighborhood weighted average: ; in The communication strength from step 1 has been normalized so that the weights sum to 1. Neighboring devices share their own [communication strength]. Values, rather than raw data, are used to protect privacy.

[0139] The fitness function is defined as the overall utility of the production line layer: ; in The target preference was initially set manually. This can be adjusted in the subsequent administrator configuration interface; The regularization coefficient is... These serve as reference values ​​for the initial weights to prevent drastic shifts. Particle position. weight Substituting this into the fitness function will give you the fitness value of the particle.

[0140] 3. Improved Quantum Behavior Particle Swarm Optimization Algorithm - Time-varying Topological Inertial Quantum Behavior Particle Swarm Optimization We employ quantum behavior particle swarm optimization as a basis, and combine time-varying topology and adaptive contraction-expansion coefficients to enhance global search capabilities. At the same time, we introduce local elite perturbations to escape local optima.

[0141] (1) Quantum behavior update equation In standard quantum swarm optimization, particle position updates depend on local attractors. and the group average optimal position This invention addresses the characteristics of the weight optimization problem by modifying the construction of the attractor and the contraction-expansion coefficient.

[0142] First, calculate the particles. Individual historical best position and the global optimal position of the entire particle swarm Define a local attractor: ; in Represents element-wise product. From A uniformly distributed sampled random vector. This operation is consistent with standard quantum behavior particle swarm optimization, but this invention further introduces neighborhood optima: a flat plate. In addition to maintaining local In addition, it receives the optimal fitness and corresponding weights of each tablet from its neighbors. The tablet with the highest fitness among all received neighborhood optima is selected and denoted as . Then, with probability Replace some dimensions of the attractor with The corresponding components enable collaborative exploration.

[0143] (2) Time-varying topological inertia and contraction-expansion coefficient In standard quantum swarm optimization, the particle position update formula is: ; in For the average optimal position, Uniformly distributed random numbers, The contraction-expansion coefficient.

[0144] This invention will Improved to adaptive time-varying coefficients, fusing information from the local search phase and the global convergence state: ; in , , The maximum number of iterations for this round of optimization is set to 50. The current global optimal fitness The average fitness of all particles, and It is an extreme value. This is an adjustment factor. When the global optimum is significantly better than the average level, Increase to enhance exploration; conversely, decrease to refine development, achieving an adaptive trade-off.

[0145] (3) Weighted projection and boundary treatment The updated particle position may violate Non-negativity constraints. Using a scaling projection method: first, clip each dimension to... Then calculate the sum. ,like Then reinitialize to a uniform distribution; otherwise execute... An implicit velocity is defined in the position difference between adjacent generations to determine convergence. If the root mean square of the position changes of all particles is less than 0.5% for five consecutive generations... If the result is positive, then the round is considered converged, and the current global optimum is retained as the result for that round.

[0146] 4. Cross-device collaborative communication mechanism Tablets communicate with each other via the HarmonyOS distributed soft bus, using a publish-subscribe model. Each tablet publishes its own... The particle swarm optimization (PSO) algorithm transmits its fitness value to a designated Topic and subscribes to the Topics of all its online neighbors. Communication occurs once after each local PSO iteration, with a single 50-generation iteration cycle completing in approximately 10 seconds. To reduce communication load, updates are only sent when the fitness improvement exceeds 1%.

[0147] Considering the instability of industrial field networks, a local cache decay mechanism is introduced: the optimal information of each neighbor is timestamped; if it exceeds [a certain value], [the cache decays]. If the neighbor does not update within a communication cycle, then the neighbor... The calculated impact weights decay linearly to 0, thus avoiding misleading optimization by outdated information.

[0148] 5. Online operation process The entire collaborative weight adjustment process runs continuously using a sliding window approach: Triggering timing: The system clock triggers this round of optimization every 30 minutes. Additionally, when a new industrial tablet is added to the production line or a device detects a sharp deterioration in recommended metrics, such as a download rate drop of more than 20%, a new round of optimization will be immediately triggered to quickly adapt.

[0149] Data collection: tablet Statistics from the local recommendation log for the previous 30-minute window ,calculate And broadcast requests to neighbors to share them. Value. Receive neighbor responses and integrate them to calculate the neighborhood average. .

[0150] Particle swarm initialization: 20 particle positions are initialized using an optimal point set to ensure uniform distribution of individuals on the 3D simplex. If an optimal weight from the previous round exists... Then initialize the position of one of the particles to The rest are randomly perturbed using chaotic mapping. Generate to leverage historical experience.

[0151] Particle swarm iteration: Execute the aforementioned improved quantum behavior particle swarm algorithm for 50 generations, evaluating fitness in each generation and updating... and It also exchanges optimal information with other tablets. After 50 generations of convergence, the optimal weight for the current round of this device is obtained. .

[0152] Weight update and smoothing: To avoid weight jitter, exponential smoothing is applied to the final output. ; Smoothing coefficient , This represents the final weights from the previous round. (Smoothed) The comprehensive score module in step 5 will be updated immediately for real-time recommendation ranking over the next 30 minutes.

[0153] Log recording and model feedback: The fitness value, final weight, and neighborhood shared data in this round of optimization are all written to the local behavior log, which will serve as incremental training material for the virtual world social simulation in the future step 4 hyperparameter search stage, forming a closed loop.

[0154] Finally, the input and output of step 6 are: enter: Statistical metrics of recommendation performance for local and neighboring devices over the past 30 minutes: and neighboring equipment value; Final weight of the previous round Initial value is taken ; The online neighbor list and communication strength maintained by HarmonyOS distributed soft bus ; Output: Smoothed optimal comprehensive weight vector for this device ; Update to step 5 for subsequent recommendation sorting.

[0155] Through the aforementioned cross-device collaborative weight adjustment mechanism based on improved quantum behavior particle swarm optimization, the recommendation strategy for each industrial panel within the production line can dynamically adapt to changes in global efficiency and local demands. This maintains the individuality of each workstation while achieving overall optimization at the production line level through swarm intelligence. This step is closely integrated with the preceding and following steps: the data comes from the inference log in step 5 and the topology information in step 1, and the output weight vector directly controls the recommendation ranking in step 5. The entire process forms a continuous online adaptive optimization closed loop.

[0156] The present invention also provides several embodiments to further illustrate the above-described method: Example 1: Performance comparison with existing recommendation methods To verify the superiority of the proposed industrial tablet AI software store system and management method based on HarmonyOS, this invention uses real-world operation logs from 30 HarmonyOS industrial tablets on a large-scale electronics manufacturing production line over six months as a dataset. The proposed method (hereinafter referred to as "this method") is then compared offline with three baseline methods: collaborative filtering, standard DCN, and multi-task MMOE. Evaluation metrics include: Download Prediction AUC: The area under the ROC curve for model prediction of download behavior; the larger the value, the better. Adaptation Prediction (MAE): The mean absolute error between the predicted adaptability score and the actual operating status label; the smaller the value, the better. Safety compliance accuracy: The accuracy of binary classification prediction for safety compliance; Overall Recommendation Hit Rate @20: The percentage of the top 20 recommendations that were actually downloaded or successfully installed by the operator.

[0157] All methods used the same training / validation / test set partitioning, with the test set containing 20,000 samples. The comparison results are shown in Table 1.

[0158] Table 1 Performance Comparison of Different Recommendation Methods

[0159] As shown in Table 1, our proposed method achieves optimal results across all four metrics. Compared to the standard DCN, the download prediction AUC is improved by approximately 9.2%, the adaptation prediction MAE is reduced by approximately 25.9%, the security compliance accuracy is improved by approximately 8.2%, and the overall hit rate @20 is improved by approximately 12.9%. This is attributed to the combined advantages of our method, including the multi-head attention adaptive cross-network, the dual-path frequency domain enhanced deep network, and the hierarchical dynamic routing expert system. Furthermore, by improving the bat algorithm to perform global hyperparameter optimization in a virtual world social simulation, the model exhibits stronger feature representation and generalization capabilities in industrial multi-task scenarios.

[0160] Example 2: Convergence Verification of Cross-Device Collaborative Weight Adjustment This embodiment demonstrates the online performance of the cross-device collaborative weight adjustment method based on improved quantum behavior particle swarm optimization in step 6. An industrial plate at a milling station in the production line is selected, and its smoothed comprehensive score weight vector is monitored over 30 consecutive optimization cycles (30 minutes per cycle). The process of change was investigated to verify the convergence of the algorithm and the rationality of the weight adjustment. Figure 2 This is the curve showing the change of the weight components of the flat plate.

[0161] Figure 2 This indicates that the weights were initially set manually. Based on this, after approximately 5 optimization cycles, the download preference weight was adjusted. It gradually rises and stabilizes in the 0.5-0.6 range, adapting to the bias weight. It has slightly decreased to around 0.25, indicating a shift in the safety bias weight. The value was fine-tuned to around 0.2. This is because the operator at this workstation focuses more on software downloads to meet the requirements of new processes during stable production periods. The high success rate and high safety rate information shared by neighboring devices mean that safety and adaptation weights do not need to be excessively high to ensure basic compliance. No drastic oscillations occurred throughout the process, and the sum of the weights remained constant at 1, verifying the effectiveness of the improved quantum behavior particle swarm optimization algorithm combined with scaling projection and exponential smoothing, as well as the guiding role of neighborhood information collaboration in achieving local optima.

[0162] This invention provides an industrial tablet software recommendation system based on swarm intelligence. This system corresponds one-to-one with the steps of the aforementioned method and is deployed on the HarmonyOS industrial tablet and its distributed production line environment. The system comprises six collaborative subsystems: a perception and topology management subsystem, a multimodal feature encoding subsystem, a swarm intelligence-enhanced multi-task deep recommendation network inference subsystem, an improved bat algorithm training and evaluation subsystem, an online security filtering engine, and a distributed collaborative weight optimization subsystem. The data flow and functions between the subsystems are as follows.

[0163] 1. Sensing and Topology Management Subsystem This subsystem corresponds to step 1 and is responsible for real-time acquisition of multi-source data from the equipment and production line environment to generate basic features. It consists of three modules: The operating condition acquisition and preprocessing module acquires multi-dimensional time-series data on the industrial tablet's operating conditions through HarmonyOS sensor services, including CPU temperature, memory usage, vibration root mean square value, and network packet loss rate. It performs outlier correction and min-max normalization, and outputs an operating condition scenario matrix. .

[0164] Production line status awareness module: Reads the current production task identifier and production line workstation role through OPC UA or Manufacturing Execution System interface, and maps them to pre-trained task embedding vectors. and role embedding vector .

[0165] Distributed topology discovery and soft bus module: Based on HarmonyOS distributed soft bus, it discovers surrounding industrial tablets of the same type and calculates communication strength based on the number of communication data packets and round-trip latency. Filter and maintain the set of neighboring devices At the same time, neighborhood information is distributed for subsequent collaboration.

[0166] The output of this subsystem is a working condition scenario matrix. Task embedding vector Role embedding vector Neighbor set and communication strength It is provided to the multimodal feature coding subsystem.

[0167] 2. Multimodal Feature Coding Subsystem Corresponding to step 2, the system receives the output from the perception and topology management subsystem, as well as the user behavior embeddings and software attribute embeddings stored in the user log submodule. It then performs deep encoding and fusion of the five modalities. This includes: The industrial operating condition feature encoding module includes a time-domain multi-scale attenuation attention channel and a frequency-domain saliency channel. The time-domain channel uses three parallel convolutional layers with kernel sizes of 7, 15, and 31 to extract multi-scale features and apply attenuation attention weighting. The frequency-domain channel performs an FFT on the four-dimensional operating condition signal, extracting the amplitude and normalized frequency of the maximum amplitude frequency component in each dimension, encoding it into an 8-dimensional vector, and then outputting it through a fully connected network. Finally, the dual-channel gated fusion outputs a 32-dimensional operating condition embedding. .

[0168] User behavior embedding module: Retrieves 32-dimensional behavior embeddings from the embedding table based on the current operator ID. For new users, a global average embedding plus compensation term is used.

[0169] Software attribute embedding module: For candidate software, it generates 32-dimensional software attribute embeddings by combining industrial BERT encoding software description text with functional tags and resource requirements. .

[0170] Distributed topology coding module: Constructs a device graph using neighbor device sets, communication strength, role embedding vectors, and operating condition embeddings, and introduces virtual nodes representing the overall status of the production line. A 32-dimensional topological embedding is obtained by executing a two-layer multi-head graph attention network. .

[0171] Gated fusion module: Calculates cross-gating coefficients for the four embeddings, modulates and concatenates them, and projects them into a 128-dimensional multimodal fusion feature vector. .

[0172] The output of this subsystem is a list of fused feature vectors of all candidate software, which is then input into the swarm intelligence-enhanced multi-task deep recommendation network inference subsystem.

[0173] 3. Swarm Intelligence Enhanced Multi-Task Deep Recommendation Network Inference Subsystem Corresponding to step 3, using the multimodal fused feature vector as input, three predicted values ​​are forward-inferred. The network structure is frozen from the optimal parameters obtained in step 4, including: Multi-head attention adaptive cross network: through 4-head attention interaction and The order controller dynamically determines the crossover order, up to 6th order. The low-rank decomposition parameter matrix, combined with layer normalization and Dropout, outputs high-order crossover features.

[0174] Dual-path deep network: The general path consists of 4 layers of residual fully connected layers, while the industrial signal enhancement path additionally fuses frequency domain features. Gating merges the outputs of the two paths.

[0175] The hierarchical dynamic routing expert system consists of 4 shared experts and 2 specific experts for each task. A two-level dynamic routing gating selects and activates the experts, and outputs the predicted embeddings for the three tasks.

[0176] Task Tower Output Layer: Each task goes through two fully connected layers, 16→1, and sigmoid activation to generate download probability. Suitability score Safety and compliance score .

[0177] This subsystem outputs a three-dimensional score for each candidate software and passes it to the online security filtering engine.

[0178] 4. Improve the training and evaluation subsystem of the bat algorithm. Corresponding to step 4, offline model training and hyperparameter optimization are completed, generating deployment models and suitability maps. This includes: Training dataset construction module: 200,000 training samples are extracted from the historical logs of multiple production lines. Each sample contains a fused feature vector generated by the multimodal feature encoding subsystem and three labels: download, adaptability, and security compliance. The training / validation / test sets are divided by time to generate a lightweight evaluation subset.

[0179] Multi-task loss calculation module: integrates homoscedastic uncertainty weighted loss, combines binary cross-entropy and mean square error, and can automatically balance three tasks by learning noise parameters.

[0180] The Chaotic Multimorphic Adaptive Bat Algorithm Module: This module initializes the population with chaotic quantum state encoding (40 bats). It employs a morphological adaptive mechanism that integrates wolf pack hierarchy, ant colony pheromone mechanisms, and fruit fly olfactory search strategies to control exploration / exploration behavior. Elite memory-guided hybrid Levy-Drosophila escape mechanisms help them break out of local optima. Fitness evaluation utilizes a virtual world social simulator to simulate the recommendation performance of a 50-device production line over one cycle, considering download rate, success rate, security rate, and model complexity regularization.

[0181] Full-data fine-tuning training module: Using the optimal solution of hyperparameters as the configuration, the AdamW optimizer and cosine annealing learning rate are used to train a swarm intelligence-enhanced multi-task deep recommendation network on the full training set, and early stopping is used to select the model parameters with the lowest validation loss. Output the final model and suitability map.

[0182] 5. Online security filtering engine Corresponding to step 5, protocol compatibility and security level verification are performed in real-time inference, the recommendation score is revised, and the final recommendation list is generated. This includes: Industrial protocol compatibility reasoning module: Performs a two-hop knowledge graph query on candidate software to verify the intersection of its driver dependency chain and the industrial protocols supported by the current device. If incompatible, the compatibility score is applied. Multiply by a penalty factor of 0.2.

[0183] SIL (Safety Integrity Level) Automatic Matching Module: Reads the required SIL level of the device and the declared SIL level of the software. If the software level is lower than required, it forces a change in the safety compliance score. Set to zero.

[0184] The comprehensive scoring, ranking, and output module utilizes the weight vectors distributed in real-time by the distributed collaborative weight optimization subsystem. Calculate the overall score After sorting, the top 20 recommended lists are displayed, and feedback logs are recorded.

[0185] 6. Distributed collaborative weight optimization subsystem Corresponding to step 6, the weight vector of the comprehensive score is continuously adjusted online, and decentralized group collaboration is achieved through an improved quantum behavior particle swarm optimization algorithm. This includes: Local metrics statistics module: Every 30-minute window, it counts the number of local recommendations, downloads, successful adaptations, and security compliances, and calculates smoothed sub-metrics.

[0186] Neighborhood data exchange and fusion module: Shares sub-indicators of each tablet through a distributed soft bus publish-subscribe mode, and calculates the neighborhood weighted average index by combining communication strength.

[0187] Quantum Behavior Particle Swarm Optimization Module: Each plate maintains 20 particles, employs a quantum attractor update equation, integrates neighborhood optima and local optima, adaptively shrinks and expands coefficients, and iterates through 50 generations to search for the optimal weights. .

[0188] Weight smoothing and application module: After smoothing the current round's optimal weight with the historical weight index, it updates the comprehensive scoring module to drive the recommendation ranking within 30 minutes.

[0189] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0190] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0191] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0192] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0193] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. An industrial tablet AI software store system based on HarmonyOS, characterized in that: include: The perception and topology management subsystem is used to collect multi-dimensional time-series data of equipment operating conditions, production line status identifiers and neighbor equipment connection relationships in real time, and generate operating condition scenario matrix, task embedding vector, role embedding vector, neighbor set and communication strength. The multimodal feature encoding subsystem receives data output from the perception and topology management subsystem, as well as user behavior embedding and software attribute embedding. It encodes working condition features through a dual-channel encoding of temporal multi-scale attenuation attention and frequency domain saliency, encodes distributed topology features through a graph attention network with virtual nodes, and performs gated fusion of each modality feature to generate a multimodal fusion feature vector. The swarm intelligence-enhanced multi-task deep recommendation network inference subsystem utilizes a network structure containing a multi-head attention adaptive cross network, a dual-path deep network, and a hierarchical dynamic routing expert system to perform forward inference on the multimodal fusion feature vector, while simultaneously outputting download probability, adaptability score, and security compliance score. An improved bat algorithm training and evaluation subsystem was developed to build a training dataset offline. The system combined a homoscedastic uncertainty weighted loss function and adopted a chaotic multimorphic adaptive bat algorithm that integrates wolf pack hierarchical division of labor, ant colony pheromone and fruit fly olfactory search to optimize network hyperparameters. The fitness was evaluated in a virtual world social simulation and finally fine-tuned to obtain model parameters and hyperparameter suitability map. An online security filtering engine is used to apply industrial protocol compatibility reasoning penalties to the adaptability score in real-time inference and to correct the security compliance score by matching the SIL level. Then, the weight vector issued by the distributed collaborative weight optimization subsystem is used to sum the three scores in a weighted manner to generate a ranked recommendation list. The distributed collaborative weight optimization subsystem is used to receive local and neighborhood recommendation effect statistics online. It iteratively searches for the optimal score weight of each device by constructing the neighborhood optimal participant attractor and using an improved quantum behavior particle swarm algorithm with adaptive shrinkage-expansion coefficients. After smoothing, the score weight is sent to the online security filtering engine.

2. A management method for an industrial tablet AI software store system based on HarmonyOS, applied to the industrial tablet AI software store system based on HarmonyOS as described in claim 1, characterized in that, include: Step 1: Collect multi-dimensional operating condition time-series data, production line task and role information, and neighbor device connection relationships of industrial tablets, and perform normalization preprocessing to form standardized input; Step 2: Extract multi-scale time-domain features and frequency-domain salient features from the working condition time-series data and perform gated fusion. Combine user behavior embedding, software attribute embedding, and distributed topological embedding based on graph attention coding, and obtain a multimodal fusion feature vector through cross-modal gated fusion. Step 3: Input the multimodal fusion feature vector into the swarm intelligence enhanced multi-task deep recommendation network. The swarm intelligence enhanced multi-task deep recommendation network includes an adaptive cross network in which the cross order is dynamically determined by the order controller and feature interaction is achieved through multi-head attention, a dual-path deep network that fuses general path and frequency domain enhanced path gating, and a hierarchical expert system that allocates shared experts and task-specific experts based on two levels of dynamic routing gating. Finally, the download probability, adaptability score and security compliance score are output. Step 4: In the offline training phase, a training sample set containing download, adaptation, and security compliance labels is constructed. With homoscedastic uncertainty weighted loss as the optimization objective, an improved bat algorithm that combines chaotic quantum state encoding and multiple swarm intelligence strategies is used to optimize the network hyperparameters. Among them, fitness evaluation is carried out through virtual world social simulation. After obtaining the optimal hyperparameters, the model parameters are fine-tuned. Step 5: During online reasoning, the adaptability score output in Step 3 is penalized by industrial protocol compatibility knowledge graph reasoning, and the security compliance score is corrected by SIL level matching. Then, the three corrected scores are weighted and summed using dynamically updated score weights, and a recommendation list is generated by sorting the scores according to the comprehensive score. Step 6: Through a distributed collaborative mechanism, each industrial tablet periodically calculates local recommendation performance indicators, shares indicators with its neighbors, and calculates a neighborhood weighted average to improve the quantum behavior particle swarm algorithm for searching local optimal score weights. After smoothing, the score weights are used to update the score weights in Step 5.

3. The industrial tablet AI software store management method based on HarmonyOS system according to claim 2, characterized in that, In step 1, the multi-dimensional operating condition time-series data includes the CPU temperature, memory usage, root mean square value of vibration, and network packet loss rate of the industrial tablet. The neighbor device connection relationship is obtained by discovering industrial tablets of the same type through the HarmonyOS distributed soft bus and calculating the communication strength.

4. The industrial tablet AI software store management method based on HarmonyOS system according to claim 2, characterized in that, In step 2, the multi-scale temporal features are obtained by applying one-dimensional convolutions with different kernel sizes to the time series data of the working conditions and then using attention-weighted aggregation with time decay. The significant feature in the frequency domain is obtained by independently performing a fast Fourier transform on each working condition dimension, extracting the amplitude and frequency of the frequency component with the largest amplitude, and encoding them. The distributed topology coding uses industrial tablets as nodes, neighbor relationships and communication strength as edges, and adds virtual nodes representing the overall state of the production line. The topology embedding is generated through a multi-head graph attention network.

5. The industrial tablet AI software store management method based on HarmonyOS system according to claim 4, characterized in that, In step 3, in the adaptive cross network, each cross layer interacts with the initial input and the output of the previous layer through multiple attention heads. The resulting multi-head outputs are spliced ​​together and fused with the cross terms. The sequence controller outputs the termination probability distribution of each order. When the cumulative termination probability exceeds the threshold, the cross is stopped. In the hierarchical expert system, the first level of the two-level dynamic routing gating is a sparse gating, and the second level introduces learnable prior attention to achieve fine-tuning of expert weights.

6. The industrial tablet AI software store management method based on HarmonyOS system according to claim 5, characterized in that, In step 4, the improved bat algorithm, which encodes chaotic quantum states and integrates multiple swarm intelligence strategies, specifically includes: Initial values ​​for quantum state angles are generated using Tent chaotic mapping; The population was divided into three layers based on fitness: alpha wolf, follower wolf, and scout wolf. Different layers correspond to different morphological types and search strategies. In the search space, a pheromone matrix is ​​maintained, and elite bats release pheromones and bias the quantum rotation angle. When the global optimum is continuously stagnant, Levy escape is performed by combining the historical position of the elites and the random perturbation of the fruit fly's sense of smell.

7. The industrial tablet AI software store management method based on HarmonyOS system according to claim 6, characterized in that, Step 4, the virtual world social simulation includes: A production line simulation environment with multiple virtual industrial tablets is constructed. The virtual working conditions are generated by a temporal generative adversarial network, and distributed collaboration between virtual devices is simulated. The fitness function is used to comprehensively measure the weighted result of recommendation download rate, installation success rate, security compliance rate and model complexity.

8. The industrial tablet AI software store management method based on HarmonyOS system according to claim 7, characterized in that, In step 5, the adaptability score output in step 3 is penalized using an industrial protocol compatibility knowledge graph reasoning method, specifically as follows: Two-hop queries are performed using a pre-built industrial knowledge graph to verify the compatibility between the software-driven dependency chain and the device communication protocol. If incompatible, a penalty coefficient is applied to the adaptability score. SIL level matching compares the required security integrity level of the device with the declared security integrity level of the software. When the software level is lower than the requirement, the security compliance score is set to zero.

9. The industrial tablet AI software store management method based on HarmonyOS system according to claim 8, characterized in that, In step 6, the improved quantum behavior particle swarm algorithm, when constructing local attractors, probabilistically replaces some dimensional components with the globally optimal position received from the neighbors. Its contraction-expansion coefficient is adaptively adjusted based on the difference between the current global optimal fitness and the average fitness of the particle swarm; Search results are output after exponential smoothing to avoid weight fluctuations.

Images