Multi-task dynamic scheduling system and method based on multi-source perception edge calculation

Abstract

The invention provides a multi-task dynamic scheduling system and method based on multi-source perception edge computing. Comprising a multi-source sensing module which is used for accessing a multi-source sensor and carrying out time base alignment on collected data so as to form a multi-source data stream; the edge calculation processing module is used for carrying out preprocessing, model reasoning and feature fusion on the multi-source data flow and outputting events and features for task generation; the task modeling module is used for generating task objects according to the events and the features; the resource arrangement module is used for distributing the tasks to a fast channel, a shared channel or a low-power-consumption channel according to a priority threshold value and deadline gating; the energy efficiency management module is used for constructing a task energy consumption model based on the working duration of the sensor and the processing load and generating an energy consumption budget; and the control execution module is used for issuing a control instruction to external equipment. According to the invention, end-side low-delay deterministic control, elastic resource arrangement and energy consumption collaborative optimization can be realized, and stable parallel of multiple tasks is supported.

Description

Technical Field

[0001] This application relates to the field of industrial automation technology, and in particular to a multi-task dynamic scheduling system and method based on multi-source sensing edge computing. Background Technology

[0002] In Industry 4.0 scenarios, industrial controllers need to simultaneously handle multi-source sensing data processing (vision, vibration, temperature, etc.), real-time control command execution, and energy consumption optimization scheduling on the edge side, requiring low latency, determinism, and high concurrency under limited computing power and memory conditions.

[0003] Existing technologies mostly employ fixed priorities or static scheduling, supplemented by coarse-grained resource isolation and simple energy-saving strategies, or move perception computing to the central cloud side. The former makes it difficult to reconfigure resources according to changes in service phases, and is prone to competition and priority inversion among central processing units, acceleration units, and storage bandwidth; the latter introduces link latency and jitter, and end-to-end timing is uncontrollable.

[0004] This leads to the following problems: the control link, which is sensitive to millisecond-level response, such as emergency stop and fault detection, has large latency and jitter (the delay of emergency stop command often exceeds 200 milliseconds); under multi-task concurrency, high-priority control tasks may be blocked by non-critical sensing tasks; the normal operation of sensors and frequency adjustment based on average load make energy consumption difficult to control, resulting in a low energy efficiency ratio; when the controller memory does not exceed 512 megabytes, the overhead of general framework and model metadata further squeezes the available resources. Summary of the Invention

[0005] In view of this, embodiments of this application provide a multi-task dynamic scheduling system and method based on multi-source sensing edge computing to solve the problems of concurrent resource contention and priority inversion, insufficient end-to-end real-time performance and time determinism of control instructions, and energy efficiency imbalance caused by lack of linkage between computing power and sampling configuration under energy consumption constraints in the existing technology.

[0006] The first aspect of this application provides a multi-task dynamic scheduling system based on multi-source sensing edge computing, comprising: a multi-source sensing module for accessing multi-source sensors and aligning the collected data to a time base to form a multi-source data stream; an edge computing processing module for preprocessing, model inference, and feature fusion of the multi-source data stream, and outputting events and features for task generation; a task modeling module for generating task objects based on events and features, and setting an attribute set for each task object including priority, deadline, energy consumption coefficient, and resource requirement parameters; and a resource orchestration module for orchestrating resources according to the attribute set for a central processing unit, a neural network processing unit, and a storage band. The system performs joint orchestration with the bus channel and distributes tasks to fast channels, shared channels, or low-power channels according to priority thresholds and deadline gating. It also performs interrupt preemption and recovery when there are time-limit risks or control triggers. The energy efficiency management module is used to build a task energy consumption model based on sensor working time and processing load, generate an energy consumption budget, and provide the energy consumption budget as a constraint to the resource orchestration module to adjust the sampling frequency, processing frequency, and power domain state. The control execution module is used to send control commands to external devices through the industrial communication interface according to the execution plan output by the resource orchestration module, and send the execution status back to the task modeling module for subsequent scheduling and control execution.

[0007] The second aspect of this application provides a multi-task dynamic scheduling method based on multi-source sensing edge computing, using the system of the first aspect. The method includes: acquiring data collected from multiple sources of sensors and aligning the collected data to a time base to form a multi-source data stream; preprocessing, model inference, and feature fusion of the multi-source data stream to generate events and features for task generation; generating task objects based on the events and features, and setting an attribute set including priority, deadline, energy consumption coefficient, and resource requirement parameters; constructing a task energy consumption model based on sensor operating time and processing load, generating task-level and system-level energy consumption budgets, and mapping the energy consumption budgets to constraints such as sampling frequency upper limit, processing frequency range, and power domain state; jointly orchestrating the central processing unit, neural network processing unit, storage bandwidth, and bus channel based on the attribute set and energy consumption budget, and performing interrupt preemption and checkpoint recovery when there is a time limit risk or control trigger; and issuing start / stop and parameter commands to external devices through an industrial communication interface based on the execution plan generated by the joint orchestration, and collecting execution status feedback to update the attribute set and energy consumption budget for subsequent scheduling and control execution.

[0008] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: The system employs a multi-source sensing module to access multiple sensors and align the collected data to form a multi-source data stream. An edge computing processing module preprocesses the multi-source data stream, performs model inference and feature fusion, and outputs events and features for task generation. A task modeling module generates task objects based on events and features, and sets a set of attributes for each task object, including priority, deadline, energy consumption coefficient, and resource requirement parameters. A resource orchestration module jointly orchestrates the central processing unit, neural network processing unit, storage bandwidth, and bus channels according to the attribute set and prioritizes them. Threshold and deadline gating distributes tasks to fast channels, shared channels, or low-power channels, and performs interrupt preemption and recovery when time-limit risks or control triggers occur. The energy efficiency management module builds a task energy consumption model based on sensor operating time and processing load, generates an energy consumption budget, and provides this budget as a constraint to the resource orchestration module to adjust sampling frequency, processing frequency, and power domain state. The control execution module issues control commands to external devices via an industrial communication interface based on the execution plan output by the resource orchestration module, and sends the execution status back to the task modeling module for subsequent scheduling and control execution. This application enables low-latency deterministic control at the edge, flexible resource orchestration, and energy consumption co-optimization, supporting stable parallel operation of multiple tasks. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the structural composition of a multi-task dynamic scheduling system based on multi-source sensing edge computing provided in an embodiment of this application; Figure 2 This is a flowchart illustrating the multi-task dynamic scheduling method based on multi-source sensing edge computing provided in this application embodiment. Detailed Implementation

[0011] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0012] Industry 4.0 requires controllers to process simultaneously: multi-source sensing data (visual / vibration / temperature), real-time control commands (equipment start / stop / parameter adjustment), and energy efficiency optimization tasks (dynamic power consumption adjustment).

[0013] Existing technical solutions typically employ fixed-priority scheduling or centralized cloud processing; however, these approaches suffer from high response latency and low local resource utilization.

[0014] Therefore, industrial controllers face three major challenges when performing multi-source sensing tasks (such as visual inspection, vibration analysis, and temperature monitoring): Resource contention conflict: Concurrency of multiple tasks leads to contention for computing resources (CPU / memory), and high-priority tasks (such as device fault detection) are blocked by low-priority tasks; Insufficient real-time performance: Traditional scheduling strategies cannot meet millisecond-level response requirements (e.g., emergency stop instruction delay exceeds 200ms). Energy efficiency imbalance: Continuous operation of multiple sensors and computing tasks leads to a surge in controller power consumption (energy efficiency ratio < 0.8 TOPS / W).

[0015] In view of the deficiencies in the existing technology, this application provides a multi-source sensing edge computing and multi-task dynamic scheduling system and method based on industrial controllers. This application aims to achieve the following objectives: realize elastic resource allocation for multi-source sensing tasks, ensuring the real-time performance of high-priority tasks (response latency ≤20ms); construct a dual-objective optimization model of energy consumption and latency, reducing system power consumption by >30%; and design a lightweight scheduling engine adapted to resource-constrained industrial controllers (memory ≤512MB).

[0016] First, the system architecture involved in this technical solution in a real-world scenario is described in general. The multi-source sensing edge computing and multi-task dynamic scheduling system based on an industrial controller involved in this application mainly includes the following components: multi-source sensors, edge computing embedding layer, dynamic scheduling engine, controller execution layer, energy efficiency optimization module, and industrial equipment.

[0017] The following describes in detail the specific framework and functions of the multi-task dynamic scheduling system based on multi-source sensing edge computing provided in this application, with reference to the accompanying drawings and specific embodiments. Figure 1 This is a schematic diagram of the structural composition of a multi-task dynamic scheduling system based on multi-source sensing edge computing provided in an embodiment of this application, as shown below. Figure 1 As shown, the multi-task dynamic scheduling system based on multi-source sensing edge computing can specifically include the following modules: The multi-source sensing module 101 is used to connect to multiple source sensors and perform time base alignment on the collected data to form a multi-source data stream; The edge computing processing module 102 is used for preprocessing multi-source data streams, model inference and feature fusion, and outputting events and features for task generation; The task modeling module 103 is used to generate task objects based on events and features, and to set a set of attributes for each task object, including priority, deadline, energy consumption coefficient and resource requirement parameters. The resource orchestration module 104 is used to jointly orchestrate the central processing unit, neural network processing unit, storage bandwidth and bus channel according to the attribute set, and to divert tasks to fast channel, shared channel or low power channel according to priority threshold and deadline gating, and to perform interrupt preemption and recovery when time limit risk or control is triggered. The energy efficiency management module 105 is used to build a task energy consumption model based on the sensor's working time and processing load, generate an energy consumption budget, and provide the energy consumption budget as a constraint to the resource orchestration module so as to adjust the sampling frequency, processing frequency and power domain state. The control execution module 106 is used to send control commands to external devices through the industrial communication interface according to the execution plan output by the resource orchestration module, and to send the execution status back to the task modeling module for subsequent scheduling and control execution.

[0018] In some embodiments, the multi-source sensing module includes a multi-protocol interface board, which provides an image acquisition interface, a vibration signal acquisition interface, and a temperature acquisition interface, and has electrical isolation and unified clock distribution functions; the multi-source sensing module performs hardware time stamping on the acquired data and sets a hierarchical ring buffer and water level control strategy to configure the buffer depth according to priority.

[0019] Specifically, the multi-source sensing module adopts a board-based structure to integrate image acquisition interfaces, vibration signal acquisition interfaces, and temperature acquisition interfaces, and sets up electrical isolation and unified clock distribution units on the board. This module first introduces a unified time base from the controller execution layer, generating a frequency-divided clock for camera triggering, analog-to-digital conversion, and serial bus communication through a temperature-compensated crystal oscillator and a phase-locked loop. A clock distributor then sends each clock channel to the camera input, vibration acquisition channel, and temperature communication transceiver. To avoid common-ground interference, the image differential signal, vibration analog signal, and temperature bus are respectively isolated at the channel level via digital isolators or isolation amplifiers, and electrostatic discharge protection and overvoltage protection circuits are also configured.

[0020] To implement hardware time stamping, the module incorporates a timestamp unit, using a counter driven by a unified time base as the reference. In the image channel, the timestamp unit latches the count value and appends it to the frame header when frame synchronization arrives; in the vibration channel, it latches the first segment count and records the sample interval when the analog-to-digital converter completes a set of sampling buffer writes; in the temperature channel, it latches the count value when the first byte of the response message arrives or when the response window opens after a command is issued. These timestamps are written to the module's local buffer along with the data for subsequent alignment and fusion.

[0021] To adapt to different service priorities, the module establishes a hierarchical circular cache and water level control strategy in local memory. The cache is divided into three categories based on priority: fast cache, shared cache, and low-power cache. Each category independently maintains read / write pointers, capacity, and water level thresholds. The fast cache stores frames and packets that meet the priority threshold or deadline, configured with a smaller capacity and higher read / write priority. The shared cache is used for data buffering of general tasks, with a medium capacity and employs a dual-condition dequeue based on arrival order and timestamp. The low-power cache is used for low-priority data, with a smaller capacity and allows the oldest data segments to be discarded even when the water level remains high. The water level control strategy generates an alarm flag when the water level reaches high and outputs a rate-limiting prompt to the resource orchestration module; it cancels the rate-limiting prompt when the water level reaches low. When the high water level in the shared cache or low-power cache continuously exceeds the threshold, the module performs downsampling or temporarily suspends data acquisition on low-priority channels according to the priority list.

[0022] To ensure a consistent data structure before entering the edge computing processing module, the module generates data descriptors while writing to the circular buffer. Each descriptor includes at least the sensor identifier, timestamp, data length, priority, deadline, and channel affiliation. Data descriptors and data entities are stored separately. The edge computing processing module performs zero-copy reading and registration based solely on the descriptors, thus reducing redundant data transfer. For image channels, the module uses frames or rows as the minimum dequeue unit; for vibration channels, it uses segments with a fixed number of samples; and for temperature channels, it uses complete messages. All three types of minimum units are aligned across sources using timestamps.

[0023] In terms of interface implementation, the image acquisition interface provides differential input and trigger input, with the trigger output after frequency division by a unified clock allocation unit; the vibration signal acquisition interface is configured with constant current source excitation and anti-aliasing filtering, and the sampling clock of the analog-to-digital converter is provided by the unified clock allocation unit; the temperature acquisition interface uses a differential transceiver and supports address mapping and timeout detection. The sampling configuration and channel priority of the above three types of interfaces are updated by parameter tables issued by the energy efficiency management module and the resource orchestration module. The parameter tables take effect in a time-division manner outside the cache read / write critical section to avoid interrupting ongoing acquisition.

[0024] In one optional implementation, the image channel writes the timestamp, frame number, and priority of each frame into a fast cache; the vibration channel forms a segment of 1024 points and writes it into a shared cache; and each message from the temperature channel enters a low-power cache. When the fast cache reaches a high water level, the module outputs a high water level identifier for the fast channel to the resource orchestration module, which then adjusts the execution time slots and channel affiliations of the central processing unit and the neural network processing unit accordingly. When the energy efficiency management module issues a new sampling frequency and processing frequency range, the module switches the sampling frequency and communication time slot at the segment boundary according to the parameter table, while maintaining the continuity of the timestamp.

[0025] Through the above structured implementation, the multi-source sensing module completes multi-protocol access, channel isolation, unified clock allocation, hardware time stamping and hierarchical ring buffer management, and forms a bidirectional association of parameters and status with the resource orchestration module and energy efficiency management module to support the subsequent preprocessing, inference and feature fusion processes.

[0026] In some embodiments, preprocessing, model inference, and feature fusion are performed on multi-source data streams to output events and features for task generation, including: An alignment window is established for multi-source data based on a unified time base, and integrity verification and data selection are performed. Perform preprocessing on the data that passes the verification. Preprocessing includes at least noise reduction, scale normalization, and format conversion. According to the type of data sub-stream, the corresponding inference model is selected and executed on the neural network processing unit and the central processing unit to obtain sub-features for each data sub-stream; The sub-features are aligned according to the alignment window, and fused features are generated according to the fusion rules, which include at least one of weight mapping, threshold gating, or temporal correlation. The fusion features are judged and events are generated according to the preset event judgment rules, and the events and fusion features are output together to the task modeling module to generate task objects.

[0027] Specifically, the edge computing processing module, based on the data descriptors and data entities output by the multi-source sensing module, executes the following steps in the order of window alignment → integrity verification and data selection → preprocessing → inference → sub-feature alignment → fusion → event determination → output, and submits the generated events and fused features to the task modeling module for registration as input to the task object.

[0028] To ensure cross-source synchronization, an alignment window is first established based on a unified time base. The alignment window is defined as the time interval centered on the target timestamp and covering image frames, vibration sampling segments, and temperature messages. The edge computing processing module performs integrity checks on the data entering the window, including frame header identification, sample count, and message length consistency checks. When there are missing or duplicate data within the same window, the data unit with the timestamp closest to the target time point is retained first according to the data selection strategy, and the reason for selection is marked for subsequent recording.

[0029] Preprocessing is performed on the verified data. The image substream undergoes denoising, scale normalization, and format conversion. Denoising uses median filtering or bilateral filtering to suppress high-frequency noise. Scale normalization unifies the input to the resolution preset by the edge computing module. Format conversion unifies the channel order and quantization range to the internal processing format. The vibration substream undergoes bandpass filtering, resampling, and amplitude normalization. Bandpass filtering highlights target frequency band components, resampling unifies different sampling rates to an internal reference sampling rate, and amplitude normalization eliminates range differences. The temperature substream undergoes jitter removal and range mapping, and discrete messages are unpacked into equally spaced time series for alignment with other substreams.

[0030] The inference vehicle is selected based on the data substream type and resource constraints. The image substream executes a convolutional model on the neural network processing unit to extract sub-features such as bounding boxes, categories, and confidence scores. The vibration substream executes a sequence model or spectral domain model on the neural network processing unit or central processing unit to extract sub-features such as spectral peak values, envelope energy, and trend indicators. The temperature substream executes a lightweight classifier or thresholding on the central processing unit to extract sub-features such as out-of-bounds markers and stability measures. To reduce migration latency, a memory mapping relationship is established based on data descriptors before inference, and the source, timestamp, and valid interval are registered using sub-feature descriptors after inference.

[0031] Each generated sub-feature is aligned according to an alignment window. The alignment rule is to select the sub-feature with the timestamp closest to the center of the window as the representative within the same window, and to save the continuity references of adjacent windows to support temporal correlation. Then, a fused feature is generated according to the fusion rules. The fusion rules include at least one of the following: weight mapping, threshold gating, and temporal correlation. Weight mapping is used to assign weighting coefficients to different sub-features according to source reliability and sampling freshness and synthesize them into a unified feature vector. Threshold gating is used to reject synthesis or downgrade to subset fusion when any sub-feature does not reach a confidence threshold. Temporal correlation is used to statistically analyze the persistence and rate of change of sub-features within several consecutive alignment windows and use them as the temporal components of the fused feature. The fusion process is completed on the central processing unit, and the fused feature is bound to the corresponding alignment window identifier.

[0032] Event determination is based on preset event determination rules to evaluate the fused features. These rules consist of feature thresholds, pattern templates, and context constraints. Feature thresholds are used for threshold comparisons of key components in the fused vector; pattern templates are used to match the registered pattern set; and context constraints are used to introduce limitations based on device status and operating parameters. Upon successful determination, an event is generated. The event log records fields such as event type, source window identifier, suggested priority, and suggested deadline, and is output along with the fused features to the task modeling module for generating task objects and setting attribute sets.

[0033] In one specific example, the alignment window length is set to 50ms. The image substream consists of camera image frames, which undergo median filtering, scale normalization, and internal format conversion upon entering the window. The vibration substream consists of acceleration segments sampled at 48kHz, which undergo 1kHz to 12kHz bandpass filtering and resampling to 24kHz, followed by amplitude normalization, upon entering the window. The temperature substream consists of periodic messages based on a serial bus, which undergo jitter reduction and range mapping upon entering the window to form an equally spaced time series. During the inference phase, the edge computing processing module performs a convolutional model on the image substream on the neural network processing unit to obtain defect candidate boxes and confidence scores, performs a sequence model on the vibration substream on the neural network processing unit to obtain spectral peak values ​​and envelope energy indices, and performs threshold discrimination on the temperature substream on the central processing unit to obtain out-of-bounds markers.

[0034] In the alignment phase, each sub-feature is represented by the window center timestamp. In the fusion phase, a combination of weight mapping and threshold gating is used to weight and synthesize the image defect confidence, vibration envelope energy, and temperature out-of-bounds markers, and the rate of change is calculated as the temporal component within three consecutive windows. In the event determination phase, the target event is matched according to the event determination rules, an event record is formed, and it is submitted to the task modeling module along with the fused features. The task modeling module generates the task object and sets the priority, deadline, energy consumption coefficient, and resource requirement parameters accordingly.

[0035] The above embodiments ensure that the reference relationships between alignment windows, preprocessing, inference carrier selection, sub-feature alignment, fusion rules and event determination rules are consistent, which facilitates parameter transfer and constraint configuration in the subsequent resource orchestration and energy efficiency management stages.

[0036] In some embodiments, task objects are generated based on events and characteristics, and each task object is given a set of attributes including priority, deadline, energy consumption coefficient, and resource requirement parameters, including: Based on preset mapping rules, the task type and context are determined from events and features, and a corresponding initial attribute set is generated. The deadline is determined based on the command type and interface constraints of the control execution module, and the deadline is written into the initial attribute set. The energy consumption coefficient is estimated based on the energy consumption model and combined with the sensor's working time and processing load, and the energy consumption coefficient is written into the initial attribute set. Resource requirement parameters are derived based on the model type and data scale output by the edge computing processing module; Based on changes in equipment risk levels reflected by the event, the approach of deadlines, and energy consumption budgets, online adjustments are made to priority, deadlines, energy consumption coefficients, and resource demand parameters.

[0037] Specifically, the task modeling module generates task objects and sets attribute sets based on the events and fusion features output by the edge computing processing module, following the process of "task type and context identification → initial attribute set generation → deadline determination → energy consumption coefficient estimation → resource requirement parameter derivation → online adjustment". The attribute set maintains a bidirectional correlation between parameters and status with the resource orchestration module and the energy efficiency management module for subsequent joint orchestration.

[0038] First, task type and context identification are performed. Based on preset mapping rules, the module maps the event's source channel, event category, and key components of the fusion features to the task type, and extracts contextual information, including alignment window identifiers, equipment condition identifiers, and interface constraints. In one implementation, the mapping rules map the event "abnormal vibration accompanied by visual confidence exceeding the threshold" to a "equipment anomaly review" task, and the event "temperature exceeding limits and in the parameter adjustment phase" to a "steady-state verification" task.

[0039] When generating the initial attribute set, the module creates a task object and writes placeholder fields for priority, deadline, energy consumption coefficient, and resource requirement parameters. The deadline is determined based on the command type and interface constraints of the control execution module: digital emergency stop commands correspond to shorter deadline intervals, fieldbus parameter writing corresponds to medium deadline intervals, and status queries correspond to longer deadline intervals; and the specific deadline value is obtained by combining interface bandwidth and queue depth.

[0040] The energy consumption coefficient is estimated based on the energy consumption model provided by the energy efficiency management module: E_task = k_sensor·t_duration + α·processing load^2. t_duration is determined by the task type and the expected processing stage sequence, while the processing load is estimated by the resolution, sampling rate, and model complexity of the fused features. k_sensor is selected according to the channel type and the current sampling strategy, and α is provided by the online calibration results from the energy efficiency management module. The module writes the calculated energy consumption coefficient into the initial attribute set.

[0041] The derivation of resource requirement parameters is based on the model type and data scale of the edge computing processing module: when the task involves convolutional models, a higher level is set for the neural network processing unit queue quota and storage bandwidth quota; when it involves sequence models and the sampling rate is high, the central processing unit computing quota is increased and a pass-through bus channel time slot is configured; when the task includes parameter distribution or status acquisition, the bus channel quota is configured according to interface constraints and a confirmation window is reserved. The data scale is given by the image resolution, vibration segment length, and message length, and the model type is determined by the inference carrier and operator set identifier.

[0042] After completing the initial attribute set, the module performs online adjustments. These adjustments consider the equipment risk level reflected by the event, the degree of deadline approach, and the energy budget: when the risk level increases, priority is raised and the deadline is shortened; when the deadline approach exceeds a threshold, priority is raised and resource requirement parameters are switched to the fast track mode; when the energy budget is tight and the task does not belong to the fast track, the sampling frequency and processing frequency in the resource requirement parameters are reduced. These adjustments are controlled by a hysteresis threshold to avoid frequent fluctuations near boundary conditions. The updated attribute set is submitted to the resource orchestration module to register the ready structure and time limit structure, and simultaneously written back to the energy efficiency management module for budget updates.

[0043] In a specific example, after the vibration channel detects a bearing abnormality and confirms it through fusion features, a "device abnormality verification" event is generated. The task modeling module creates a task object based on this, with the initial priority set to 8 and the deadline set to 20ms based on the emergency stop command category of the digital interface. The energy consumption coefficient is calculated based on the current vibration segment length and the image verification resolution. The k_sensor is taken as the combined value of the vibration channel and the image channel. The processing load is estimated by superimposing the complexity of the sequence model and the convolution model. The resource requirement parameters are set as high priority queue of the neural network processing unit, isolated execution unit of the central processing unit, storage bandwidth quota and pass-through bus channel time slot.

[0044] If the energy efficiency management module reports a tight budget and the control execution module does not trigger an emergency stop, the module will reduce the image verification resolution and sampling frequency for non-critical verification stages and correspondingly reduce resource requirement parameters. If the control execution module triggers an emergency stop, the module will maintain the current priority and deadline and lock the fast channel mode. After the triggering conditions are lifted, it will restore to the shared channel or low-power channel mode according to the hysteresis strategy.

[0045] Through the implementation of the above embodiments, the task modeling module completes the generation and dynamic maintenance of task objects and attribute sets from events and features, and provides a consistent parameter basis for the resource orchestration module and the energy efficiency management module.

[0046] In some embodiments, the resource orchestration module adopts a dual-gating joint orchestration mechanism based on priority and deadline to implement integrated quota control for the central processing unit, neural network processing unit, storage bandwidth and bus channels. Specifically, execution and bandwidth quotas are preset for fast channels, fair scheduling based on deadline is adopted for shared channels, cross-channel switching is achieved by combining checkpoints and interrupt preemption, and the execution frequency and sampling frequency of low-priority tasks are downgraded under the energy consumption budget constraints given by the energy efficiency management module.

[0047] Specifically, the resource orchestration module implements integrated quota control for the central processing unit, neural network processing unit, storage bandwidth, and bus channels around a dual-gating joint orchestration mechanism. The dual gating consists of a priority threshold gate and a deadline gate: when the priority in the attribute set submitted by the task modeling module meets the high threshold or the deadline is shorter than the threshold, the task is assigned to the fast channel; when the deadline is within a set range, it is assigned to the shared channel; and when the priority is lower than the lower limit and the deadline is lenient, it is assigned to the low-power channel.

[0048] The resource orchestration module establishes a channel-level quota table and scheduling structure. The quota table sets reserved entries for fast channels, allocable entries for shared channels, and baseline entries for low-power channels for the CPU execution unit queue, neural network processing unit queue, storage bandwidth, and bus channel time slots, respectively. The scheduling structure includes a ready structure sorted by priority and a time-limited structure sorted by deadline. Fast channels are bound to independent high-priority queues of CPU execution units and neural network processing units, and fixed quotas are reserved for them on the storage bandwidth and bus channel sides. Shared channels adopt a fair strategy weighted by deadlines, allocating quotas according to the deadline weights of tasks. Low-power channels perform batch processing or delayed execution based on the baseline entries.

[0049] To support cross-channel switching, the resource orchestration module is configured with checkpoints and interrupt preemption mechanisms. Checkpoints record the execution context of a task during the edge computing processing phase, including at least the sub-feature processing location, model running status identifier, buffer read / write pointers, and bus transaction sequence number. When the measurement and feedback module reports time limit risks or the control execution module triggers a control event, the resource orchestration module issues an interrupt request to elevate the preemption level of the relevant task, freezes the checkpoint of the preempted task, and switches the reserved quotas of the central processing unit execution unit, neural network processing unit queue, storage bandwidth, and bus channel to the fast channel task. After the triggering condition is lifted, the preempted task is restored based on the checkpoints, and its channel affiliation is re-determined according to dual gating.

[0050] Under the energy consumption budget constraints provided by the energy management module, the resource orchestration module implements down-level configuration for tasks in low-power channels and shared channels. The constraints consist of an upper limit for the sampling frequency, a range of processing frequencies, and power domain status. The resource orchestration module binds the constraint table to the channel quota: it performs downsampling and lowers the processing frequency for low-power channel tasks, allocates quotas for shared channels according to the budget using a weighted average, and maintains reserved quotas for fast channels while limiting the lower bound of the processing frequency for non-critical segments. Constraint updates take effect at segment boundaries based on hysteresis thresholds and smoothing windows.

[0051] In a specific example, vibration anomalies output by the multi-source sensing module and visual features together form an event. The task modeling module generates a task object and sets its priority to 8 and its deadline to 20ms. Based on this, the resource orchestration module assigns the task to the fast channel, allocates high-priority queues for the independent central processing unit execution unit and the neural network processing unit, and binds it to the reserved quota of storage bandwidth and bus channel. At the same time, the image inspection and temperature inspection tasks are assigned to the shared channel, and the remaining quota is allocated according to the weighted average of the deadline. After the energy management module issues a budget shortage indicator, the resource orchestration module performs downsampling on the temperature inspection task in the low-power channel, downsamples the processing frequency of the image inspection task in the shared channel, and keeps the reserved quota of the fast channel unchanged. When the control execution module triggers a control event, the resource orchestration module immediately switches the available quota of the shared channel and the low-power channel to the fast channel task through interrupt preemption. After the event ends, the preempted task is restored according to the checkpoint and its channel affiliation is restored according to the dual gating mechanism.

[0052] In some embodiments, tasks are routed to fast channels, shared channels, or low-power channels according to priority thresholds and deadline gating, and interrupt preemption and recovery are performed when time-limit risks or control triggers occur, including: Tasks that meet the priority threshold or have a deadline shorter than the threshold are assigned to the fast channel and bound to a reserved quota. Tasks within a set range are assigned to the shared channel and scheduled according to their deadlines. Tasks below the lower priority limit are assigned to the low-power channel and configured to be downgraded. When a time-limit risk or control trigger is detected, an interrupt request is issued to elevate the preemption level of the relevant task, the checkpoint is saved, and the quotas of the central processing unit, neural network processing unit, storage bandwidth and bus channel are switched to fast-track tasks. After the triggering conditions are lifted, the preempted task is restored based on the checkpoint, and the channel ownership is re-determined based on the dual gating mechanism.

[0053] Specifically, the resource orchestration module pre-configures priority upper limit thresholds, priority lower limit thresholds, fast channel deadline thresholds, and shared channel deadline intervals, and establishes a channel-level quota table and scheduling structure. The channel-level quota table corresponds to the reserved items, allocable items, and baseline items for the central processing unit execution unit, neural network processing unit queues, storage bandwidth, and bus channel time slots, respectively; the scheduling structure includes a ready structure sorted by priority and a time-limited structure sorted by deadline.

[0054] After the task modeling module submits the attribute set, the resource orchestration module determines channel affiliation and binds quotas based on a dual gating mechanism: tasks that meet the priority upper limit threshold or have a deadline shorter than the fast channel threshold are assigned to the fast channel and bound with reserved quotas; tasks whose deadlines fall within the shared channel range are assigned to the shared channel and registered for fair scheduling weighted by deadline; tasks with priorities lower than the lower limit and lenient deadlines are assigned to the low-power channel and loaded with a downgraded configuration strategy.

[0055] The energy consumption budget issued by the energy efficiency management module is mapped to constraints such as the upper limit of the sampling frequency, the range of processing frequency points, and the power domain state. The resource orchestration module binds it to the above-mentioned channel quotas, so that the shared channels and low-power channels are subject to budget constraints when allocating resources, while the fast channels remain reserved and the lower limit of the processing frequency point is limited in non-critical segments.

[0056] The resource orchestration module configures checkpoints and interrupt preemption mechanisms to support cross-channel switching. Checkpoints record the execution context of a task during the edge computing processing phase, including at least the sub-feature processing location, model running status identifier, buffer read / write pointers, and bus transaction sequence number. When the scheduling clock detects that the earliest deadline is about to arrive in the time-limit structure or when the control execution module triggers a control event, the resource orchestration module issues an interrupt request to elevate the preemption level of the relevant task, freezes the checkpoint of the preempted task, and immediately switches the available quota of the central processing unit execution unit, neural network processing unit queue, storage bandwidth, and bus channel to the fast channel task. After the switch is completed, the fast channel task runs along the predetermined execution sequence and occupies the reserved quota, while non-fast channel tasks remain suspended and wait for resumption.

[0057] When the triggering condition is removed, the resource orchestration module resumes operation based on the checkpoints of the preempted task, resetting its read / write pointers, model state, and bus transaction sequence number to the frozen position. It then re-determines the channel affiliation for the task through a dual-gating mechanism: if the priority has been reduced or the deadline is more lenient, it reverts to a shared channel or a low-power channel; if the fast channel conditions are still met, it continues execution within the fast channel until the current critical segment is completed. During recovery, a hysteresis threshold and segment boundary switching strategy are followed to avoid frequent back-and-forth switching near boundary conditions, which would incur additional overhead.

[0058] In one specific example, the event formed by the fusion of vibration and visual features is identified by the task modeling module as a "device anomaly review" type, with a priority set to 8 and a deadline of 20 milliseconds. The resource orchestration module assigns it to the fast channel and binds it to the high-priority queue of the central processing unit isolated execution unit and the neural network processing unit, as well as the corresponding storage bandwidth and bus channel reserved quota. The image inspection task has a priority of 4 and a deadline of 200 milliseconds, and is assigned to the shared channel, with allocable quotas weighted according to the deadline. The temperature inspection task has a priority of 2 and a lenient deadline, and is assigned to the low-power channel and loaded with downsampling and frequency reduction strategies.

[0059] During operation, if the time-limited structure detects that a fast-track task is nearing its deadline, the resource orchestration module triggers an interrupt preemption, freezing the checkpoints of tasks currently executing on the shared and low-power channels and switching available quotas to the fast-track task. After the anomaly is resolved, the resource orchestration module restores the preempted tasks based on the checkpoints and, according to a dual-gating mechanism, returns image inspection to the shared channel and maintains temperature inspection on the low-power channel. If the energy efficiency management module simultaneously issues a budget-constrained flag, the image inspection on the shared channel will have its processing frequency lowered according to the constraint table upon restoration, while the temperature inspection on the low-power channel will maintain its downsampling configuration, and the reserved quota and lower bound frequency of the fast-track task will remain unchanged.

[0060] Through the above embodiments, the resource orchestration module completes channel splitting, quota binding, and cross-channel preemption recovery under the dual gating mechanism, and links with energy consumption budget constraints to form integrated control over the central processing unit, neural network processing unit, storage bandwidth, and bus channels. This enables the continuous execution of critical tasks in concurrent scenarios, suppresses cascading blockages caused by resource contention, and reduces interference from non-critical tasks to critical segments.

[0061] In some embodiments, a task energy consumption model is constructed based on sensor operating time and processing load to generate an energy consumption budget, including: A parameterized energy consumption model is established, consisting of a sensor activity duration term and a processing load nonlinear term. The nonlinear term is characterized by a quadratic weighting of the processing load. The model parameters are calibrated and updated online based on the measurement information provided by the edge computing processing module and the control execution module; Based on the task energy consumption model, task-level and system-level energy consumption budgets are generated, and the energy consumption budgets are provided as constraints to the resource orchestration module and the task modeling module.

[0062] Specifically, the energy efficiency management module operates around the process of "metric acquisition → parametric modeling → calibration and online update → budget generation → constraint issuance". The module obtains metrics such as sensor activity duration, central processing unit and neural network processing unit utilization, storage bandwidth usage, and channel status from the edge computing processing module and control execution module. After aggregating these metrics using a sliding window, the module inputs them into the energy consumption model to generate task-level and system-level energy consumption budgets. The budgets are then mapped into constraint tables for sampling frequency upper limits, processing frequency ranges, and power domain status, and issued to the resource orchestration module and task modeling module for coordinated execution.

[0063] In some optional examples, to unify the processing load from different sources, this embodiment normalizes and weights the central processing unit occupancy, neural network processing unit utilization, and storage bandwidth occupancy to synthesize a load index L; a parameterized energy consumption model is established: E_task = k_sensor·t_duration + α·L^2, where k_sensor represents the power consumption coefficient per unit time of the sensors and peripherals involved in the current task, t_duration is the activity duration of the task in the current scheduling cycle, and α is the nonlinear weight of the processing load. In the initial calibration phase, initial values ​​of k_sensor and α are provided by offline testing; in the online update phase, recursive estimation is used to correct the parameters, with the update cycle consistent with the sliding window, and a hysteresis threshold is set to avoid frequent jitter. The system-level budget is obtained by aggregating the budgets of each task within the same cycle, while reserving a safety margin for fast channel occupancy.

[0064] In a specific example, vibration anomalies trigger visual verification. For the "Equipment Anomaly Verification" task, the energy efficiency management module calculates t_duration=40ms based on the vibration segment length and visual verification resolution, selects k_sensor (vibration + camera interface) according to the channel type, and updates α based on recent measurements; the load synthesis within the measurement window yields L=0.62. Substituting these values ​​into the model, the energy consumption budget B_task for this task is obtained, and a constraint table is formed accordingly: the fast channel maintains a high priority queue and reserved bandwidth for the neural network processing unit, limiting the lower bound of the processing frequency for non-critical segments; the shared channel image inspection lowers the processing frequency by one level and adjusts the upper limit of the sampling frequency from 30fps to 15fps; the low-power channel temperature inspection downsamples and allows the shutdown of irrelevant temperature sensors when the budget is continuously tight, maintaining state synchronization via message heartbeats during shutdown. After receiving the constraint table, the resource orchestration module binds the above constraints to the channel quota for execution; the task modeling module references the system-level budget when generating attribute sets for other tasks, providing priority downgrading suggestions for non-fast channel tasks.

[0065] Therefore, this embodiment implements an energy consumption model that combines the sensor activity duration term with the processing load nonlinear term and a two-layer budget mechanism into an executable sampling and frequency point constraint, realizing a closed-loop linkage from budget to quota to channel strategy. This ensures that in abnormal review scenarios, the fast channel maintains its predetermined quota while the non-fast channel is controlled to be downgraded according to the budget, reducing the unnecessary continuous operation of sensors and high-frequency processing on the overall resources.

[0066] In some embodiments, an energy consumption budget is provided as a constraint to the resource orchestration module to adjust the sampling frequency, processing frequency, and power domain state, including: The energy consumption budget is mapped to a constraint table of sampling frequency upper limit, processing frequency range and power domain state, and bound to the quotas of central processing unit, neural network processing unit, storage bandwidth and bus channel. Based on the constraint table, perform downsampling and frequency reduction on low-power channel tasks, allocate quotas for shared channels according to budget weighting, and maintain reserved quotas for fast channels while limiting the lower bound of processing frequency points for non-critical segments; The adjustment of sampling frequency, processing frequency and power domain state is stabilized by using a smooth window and hysteresis threshold, and takes effect in the predetermined switching order when the budget is updated.

[0067] Specifically, after generating the energy consumption budget, the energy efficiency management module maps the budget to a constraint table consisting of a sampling frequency upper limit, processing frequency range, and power domain status, along with control fields such as task identifier, channel affiliation, non-critical segment markers, smoothing window length, and hysteresis threshold. Upon receiving the constraint table, the resource orchestration module binds it to the channel-level quota table: configuring the processing frequency range for the execution units of the central processing unit, configuring the frequency and power domain control of high and low priority queues for the neural network processing units, configuring the token rate upper limit for storage bandwidth, and configuring the time slot percentage and burst threshold for bus channels; the binding relationship takes effect at the segment boundary and records the effective timestamp.

[0068] The resource orchestration module implements differentiated configurations for different channels based on the constraint table. For high-speed channels, the reserved quota remains unchanged, and the lower bound of the processing frequency is limited in non-critical sections to avoid excessive frequency reduction. Critical sections maintain the predetermined frequency and always-on power domain. For shared channels, the allocable quotas of the central processing unit, neural network processing unit, storage bandwidth, and bus channels are weighted according to the budget, and the processing frequency is limited within the constraint range. For low-power channels, downsampling and frequency reduction are prioritized. When the hysteresis threshold of budget constraints is continuously reached, the power domain of the relevant sensor or processing unit is allowed to switch to sleep mode. Sleep mode maintains message heartbeat and status word updates.

[0069] To avoid frequent jitter, the resource orchestration module employs tiered smoothing and hysteresis control for adjusting sampling frequency, processing frequency, and power domain status. Tiered smoothing aggregates metrics from multiple scheduling cycles within a smoothing window before issuing adjustment commands. Hysteresis control sets upper and lower thresholds for entry and exit conditions, and executes sampling frequency adjustment, processing frequency adjustment, and power domain switching sequentially according to a predetermined switching order. The next adjustment can only be executed after the previous one has stabilized. Power domain switching only takes effect at segment boundaries and when a continuity threshold is met.

[0070] In one specific example, the "device anomaly review" task triggered by vibration and visual fusion features is assigned to the fast channel, maintaining the high priority queue and storage bandwidth reservation of the neural network processing unit unchanged, and limiting the processing frequency of non-critical segments to no less than the third level; the image inspection task is in the shared channel, with the constraint table adjusting the upper limit of the sampling frequency from 30 frames per second to 15 frames per second, and limiting the processing frequency of the central processing unit and the neural network processing unit to between the second and third levels, while the storage bandwidth token rate is reduced according to the budget weight; the temperature inspection task is in the low-power channel, first halving the sampling frequency, and when the budget tightness indicator remains true for three consecutive smoothing windows, the temperature acquisition power domain is switched to sleep mode, retaining only the message heartbeat. After the budget is released, the processing frequency and sampling frequency are restored sequentially according to the hysteresis threshold and switching order, and the power domain is restored to normal operation after the condition is continuously released for two smoothing windows. During this period, all switching is performed at the segment boundary and the effective timestamp is recorded.

[0071] Through the above embodiments, the energy consumption budget is structured into executable sampling frequency, processing frequency and power domain state constraints and is bound to the channel quota, realizing closed-loop linkage between budget constraints and resource orchestration; without changing the reserved resources of the fast channel, the shared channel and low power channel are subject to controlled downgrading and orderly recovery, reducing unnecessary sampling and high-frequency processing, and suppressing jitter and resource waste caused by frequent switching.

[0072] In some embodiments, the system further includes: The symbol table storage module is used to store and load model weights, operators, and graph structure metadata in a lightweight symbol table. The symbol table storage module adopts relocatable indexes and variable-length encoding and supports constant pool sharing and differential encoding to provide on-demand dereferencing for lightweight container runtime submodules.

[0073] Specifically, the symbol table storage module adopts a segmented lightweight layout, including a version and check segment, a symbol directory segment, a graph structure segment, a constant pool segment, and a string pool segment. Each segment is arranged with contiguous addresses in non-volatile storage and mapped to a contiguous memory region at once during loading. The version and check segment records the version number, segment boundary descriptor, and integrity check code. The symbol directory segment establishes symbol directories for operators, tensors, parameter sets, and subgraphs, with directory entries represented by triples of symbol identifier, relocatable index, and attribute flag. The graph structure segment represents the computation graph using a separate node table and edge table. The node table records node types and input / output indices, while the edge table records topological connections. The constant pool segment stores immutable data such as weight blocks, quantization scales, and lookup tables. The string pool segment stores symbol names and alias hashes for deduplication and retrieval.

[0074] To reduce pointer overhead and support cross-address space sharing, the relocatable index uses a relative offset to represent the starting position of the referenced object within its segment. During loading, a one-time relocation is performed using the mapped base address plus the relative offset. The relative offset is represented using variable-length encoding, where the high-order bits serve as a continuation flag and the low-order bits as valid numerical bits, with the integer value obtained through byte-by-byte parsing. Differential encoding is used for edge tables and sequentially adjacent directory entries, recording only the difference between adjacent indices to reduce redundancy. To further reduce redundant storage, the constant pool segment performs hash merging on content-equivalent weight blocks, quantization scales, and lookup tables. The merged constant object is retained only once in the symbol directory segment, and a reference count is established for multi-model sharing.

[0075] In the loading process, the loader of the symbol table storage module first maps each segment to contiguous memory and verifies its version and integrity. Then, it parses the symbol directory segment to generate a fast retrieval structure from symbols to relocatable indexes. The parsed graph structure segment establishes a view of the relative relationships between nodes and edges without expanding the actual pointers. When the lightweight container running submodule requests operators or parameters, the loader retrieves the relocatable index based on the symbol identifier, calculates the memory address and length of the target object, and returns a read-only view handle to achieve on-demand dereferencing. For scenarios that require switching models or replacing subgraphs, the module uses a double-buffering strategy to complete the mapping and verification of the new symbol table in the background. When the container is at a segment boundary, it atomically switches to the new view at the handle level, and the old view is released at a safe point based on the constant pool reference count.

[0076] Regarding concurrency and alignment strategies, contiguous memory regions are organized with little-endian alignment, directory entries are aligned to 2-byte or 4-byte boundaries, and the constant pool is aligned to 16-byte boundaries for direct reading by neural network processing units. Read-only views that are dereferenced on demand can be shared concurrently by multiple container instances. Write requests are placed in the container's private buffer using a copy-on-write strategy, without being written back to the symbol table area. To ensure stable field semantics, the string pool establishes hash indexes for symbol names and aliases, allowing the container to locate the same symbol directory entry via alias retrieval. When there are shared preprocessing parameters across models, the symbol directory creates an independent entry for that shared object and references that entry in the multi-model directory.

[0077] In a specific example, the edge computing processing module needs to load both a convolutional model and a sequence model simultaneously. The symbol table storage module establishes node tables and edge tables for the graph structure segments of both models, but in the constant pool segment, the identical preprocessing lookup tables and quantization scales for both models are hashed and merged, retaining only one constant object. The convolutional kernel weight blocks in the convolutional model and the state transition matrix in the sequence model are recorded using relative offset + variable-length encoding, respectively. The edge table of the convolutional model uses differential encoding to represent the connection relationships between adjacent nodes.

[0078] After the loader completes the one-time mapping, the lightweight container runtime submodule dereferences the corresponding convolutional kernel weight view as needed during the feature extraction phase of the convolutional model by using symbolic identifiers, and dereferences the state transition matrix view as needed during the temporal inference phase of the sequence model. When it is necessary to replace a subgraph of the convolutional model online, the loader maps a new graph structure segment and constant pool segment in the background. The replacement is completed by handle switching at the safe point where the segment to be processed ends, and the constant objects of the original segment are released after the reference count reaches zero.

[0079] Through the above embodiments, the symbol table storage module achieves compact storage and on-demand loading of model weights, operators and graph data through a combination of relocatable indexes, variable-length encoding, constant pool sharing and differential encoding. The on-demand dereference mechanism between the loader and the lightweight container running submodule enables multiple models and multiple instances to share read-only views and supports atomic hot switching, reducing metadata usage and shortening loading and switching paths under limited memory conditions.

[0080] The above embodiments have described in detail the specific modules and functions of the multi-task dynamic scheduling system based on multi-source sensing edge computing of this application. The implementation process of the multi-task dynamic scheduling method based on multi-source sensing edge computing of this application will be described in detail below with reference to specific embodiments. Figure 2 This is a flowchart illustrating the multi-task dynamic scheduling method based on multi-source sensing edge computing provided in an embodiment of this application, as shown below. Figure 2 As shown, the multi-task dynamic scheduling method based on multi-source sensing edge computing can specifically include the following steps: S201: Acquire data from multiple sensors and perform time-base alignment on the acquired data to form a multi-source data stream; S202 performs preprocessing, model inference, and feature fusion on multi-source data streams to generate events and features for task generation. S203, Generate a task object based on events and characteristics, and set a set of attributes including priority, deadline, energy consumption coefficient and resource requirement parameters; S204 constructs a task energy consumption model based on sensor working time and processing load, generates task-level and system-level energy consumption budgets, and maps the energy consumption budgets to constraints such as upper limit of sampling frequency, processing frequency range and power domain state. S205, based on the attribute set and energy consumption budget, jointly orchestrates the central processing unit, neural network processing unit, storage bandwidth and bus channel, and performs interrupt preemption and checkpoint recovery when time limit risks or control triggers. S206, based on the execution plan generated by joint orchestration, sends start / stop and parameter commands to external devices through the industrial communication interface, and collects execution status feedback to update the attribute set and energy consumption budget for subsequent scheduling and control execution.

[0081] Specifically, in the multi-source sensing module, image, vibration, and temperature channels are connected via multi-protocol interface boards. A unified time base is generated by a temperature-compensated crystal oscillator and a phase-locked loop, and sent to the camera trigger, analog-to-digital converter, and serial transceiver via a clock distribution unit. Each channel has a hardware time stamp attached at the acquisition end, and the data is written to a hierarchical circular buffer according to channel priority. At the same time, a data descriptor is generated, including sensor identifier, timestamp, data length, priority, cutoff time, and channel affiliation. The edge computing processing module uses this information to establish an alignment window, organizing the smallest data units within the same window into a multi-source data stream.

[0082] Further, preprocessing is performed on the data that enters the alignment window and passes the integrity check: the image substream undergoes denoising, scale normalization, and internal format conversion; the vibration substream undergoes bandpass filtering, resampling, and amplitude normalization; and the temperature substream undergoes jitter reduction, range mapping, and discretization into equally spaced sequences. Then, an inference carrier and model are selected according to the substream type. Convolutional and sequence inference are performed on the neural network processing unit, and lightweight discrimination is performed on the central processing unit to obtain sub-features, which are then registered with their source and timestamp using sub-feature descriptors. After time alignment of the sub-features within the alignment window, fused features are generated according to fusion rules, which include at least one of weight mapping, threshold gating, or temporal correlation. Finally, the fused features are judged according to event determination rules, generating event records containing event type, source window identifier, suggestion priority, and suggestion deadline, which are output along with the fused features.

[0083] Furthermore, the task modeling module determines the task type and context based on the mapping relationship between events and fused features, creates task objects, and generates an initial attribute set. The deadline is determined based on the command category and interface constraints of the control execution module and written into the set; the energy consumption coefficient is estimated based on the energy consumption model of the energy efficiency management module, combined with sensor operating time and processing load, and then written into the set; resource requirement parameters are derived based on the model type and data scale of the edge computing processing module, including at least the central processing unit computing quota, neural network processing unit queue quota, storage bandwidth quota, and bus channel quota. Subsequently, the priority, deadline, energy consumption coefficient, and resource requirement parameters are adjusted online based on changes in equipment risk level, the approach of the deadline, and the energy budget, and submitted to the resource orchestration module to register the ready structure and time limit structure.

[0084] Furthermore, the energy efficiency management module establishes a parameterized energy consumption model, E_task = k_sensor·t_duration + α·load squared, consisting of sensor activity duration and processing load nonlinearity. The model parameters are calibrated and updated online using metrics provided by the edge computing processing module and the control execution module. Based on this model, task-level and system-level energy consumption budgets are obtained. These budgets are mapped to constraint tables defining sampling frequency limits, processing frequency ranges, and power domain states, along with task identifiers, channel affixing, and activation policies, and are then distributed. These constraint tables serve as common constraints for resource orchestration and task modeling, taking effect at segment boundaries using smoothing windows and hysteresis thresholds.

[0085] Furthermore, the resource orchestration module implements integrated quota control for the central processing unit, neural network processing unit, storage bandwidth, and bus channels under energy consumption budget constraints. A dual-gating mechanism based on priority thresholds and deadline gating is used to determine channel allocation: tasks meeting high priority thresholds or deadlines shorter than the threshold are assigned to the fast channel and bound to reserved quotas; tasks with deadlines within a set range are assigned to the shared channel and scheduled fairly based on deadline weights; low-priority tasks with lenient deadlines are assigned to the low-power channel and subjected to a downgrading strategy. To support cross-channel switching, a checkpoint and interrupt preemption mechanism is implemented: when a time-limit risk or control trigger is detected, the preemption level of the relevant task is increased, the checkpoint of the preempted task is frozen, and the available quota is switched to the fast channel task; after the trigger condition is removed, the checkpoint is restored, and channel allocation is re-determined according to the dual-gating mechanism.

[0086] Furthermore, based on the execution plan generated by the joint orchestration, the control execution module issues start / stop and parameter commands through the industrial communication interface. The commands are submitted atomically and their sequence numbers and timestamps are recorded. Execution status and fault information are fed back from field devices and written into the measurement and feedback path. Simultaneously, the attribute set in the task modeling module and the budget status in the energy efficiency management module are updated for online adjustment and constraint refresh in the next scheduling cycle, forming a closed loop of acquisition → processing → modeling → budgeting → orchestration → execution → feedback.

[0087] For example, in a bearing vibration anomaly scenario, step S201 establishes an alignment window with a length of 50 milliseconds, with image frames, vibration segments, and temperature reports entering the window according to timestamps. Step S202 performs median filtering and scale normalization on the image, bandpass filtering and resampling on the vibration from 1 kHz to 12 kHz, and de-jittering and range mapping on the temperature. The image convolution model and vibration sequence model inference are completed on the neural network processing unit. The fusion rule uses a combination of weight mapping and threshold gating to generate fusion features and determines the "equipment anomaly review" event. Step S203 generates a task object, sets its priority to 8 and its deadline to 20 milliseconds, derives resource requirements based on model complexity and data scale, and writes them into a set. Step S204 calculates the energy consumption coefficient and task budget based on the near-window metric, forms a constraint table, and distributes it. Step S205 assigns the task to the fast lane, binding the high-priority queues of the central processing unit's isolated execution unit and the neural network processing unit, along with the corresponding storage bandwidth and bus channel reserved quotas. When a time-limit risk is detected, an interrupt preemption is triggered, and the shared channel and low-power channel tasks are restored according to the checkpoint after the event is resolved. Step S206 issues emergency stop and parameter reset commands from the control execution module, and the returned execution status is used to correct the attribute set and budget for the next cycle. The above process ensures that the alignment window, events and features, attribute set, energy consumption budget, and channel quota are continuously referenced between steps, facilitating the execution of method-level executable implementations under resource-constrained conditions.

[0088] It should be understood that the sequence number of each step in the above method embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0089] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although the technical solutions of this application have been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A multi-task dynamic scheduling system based on multi-source sensing edge computing, characterized in that, include: The multi-source sensing module is used to connect to multiple source sensors and perform time base alignment on the collected data to form a multi-source data stream; The edge computing processing module is used for preprocessing, model inference and feature fusion of the multi-source data stream, and outputting events and features for task generation; The task modeling module is used to generate task objects based on the events and features, and to set a set of attributes for each task object, including priority, deadline, energy consumption coefficient and resource requirement parameters. The resource orchestration module is used to jointly orchestrate the central processing unit, neural network processing unit, storage bandwidth and bus channel according to the attribute set, and to divert tasks to fast channels, shared channels or low-power channels according to priority thresholds and deadline gating, and to perform interrupt preemption and recovery when time limit risks or control triggers occur. The energy efficiency management module is used to build a task energy consumption model based on sensor working time and processing load, generate an energy consumption budget, and provide the energy consumption budget as a constraint to the resource orchestration module so as to adjust the sampling frequency, processing frequency and power domain state. The control execution module is used to send control commands to external devices through the industrial communication interface according to the execution plan output by the resource orchestration module, and to send the execution status back to the task modeling module for subsequent scheduling and control execution.

2. The system according to claim 1, characterized in that, The multi-source sensing module includes a multi-protocol interface board, which provides an image acquisition interface, a vibration signal acquisition interface, and a temperature acquisition interface, and has electrical isolation and unified clock distribution functions. The multi-source sensing module performs hardware time stamping on the acquired data and sets up a hierarchical ring buffer and water level control strategy to configure the buffer depth according to priority.

3. The system according to claim 1, characterized in that, The preprocessing, model inference, and feature fusion of the multi-source data streams, outputting events and features for task generation, include: An alignment window is established for the multi-source data based on a unified time base, and integrity verification and data selection are performed. Preprocessing is performed on the data that passes the verification, and the preprocessing includes at least noise reduction, scale normalization and format conversion; According to the type of data sub-stream, the corresponding inference model is selected and executed on the neural network processing unit and the central processing unit to obtain sub-features for each data sub-stream; The sub-features are aligned according to the alignment window, and fused features are generated according to the fusion rules, which include at least one of weight mapping, threshold gating, or temporal correlation. The fusion feature is judged and an event is generated according to the preset event judgment rules. The event and the fusion feature are then output to the task modeling module to generate a task object.

4. The system according to claim 1, characterized in that, The step of generating task objects based on the events and characteristics, and setting a set of attributes for each task object including priority, deadline, energy consumption coefficient, and resource requirement parameters, includes: Based on preset mapping rules, the task type and context are determined from the events and features, and a corresponding initial attribute set is generated; The deadline is determined based on the command category and interface constraints of the control execution module, and the deadline is written into the initial attribute set; The energy consumption coefficient is estimated based on the energy consumption model and combined with the sensor's working time and processing load, and the energy consumption coefficient is written into the initial attribute set; Resource requirement parameters are derived based on the model type and data scale output by the edge computing processing module; Based on the changes in equipment risk level reflected by the event, the degree of approach to the deadline, and the energy consumption budget, the priority, deadline, energy consumption coefficient, and resource demand parameters are adjusted online.

5. The system according to claim 1, characterized in that, The resource orchestration module adopts a dual-gating joint orchestration mechanism based on priority and deadline to implement integrated quota control for the central processing unit, neural network processing unit, storage bandwidth and bus channels. Specifically, it presets execution and bandwidth quotas for fast channels, adopts fair scheduling based on deadline for shared channels, and achieves cross-channel switching by combining checkpoints and interrupt preemption. Under the energy consumption budget constraints given by the energy efficiency management module, it downgrades the execution frequency and sampling frequency of low-priority tasks.

6. The system according to claim 1, characterized in that, The process of routing tasks to fast channels, shared channels, or low-power channels based on priority thresholds and deadline gating, and performing interrupt preemption and recovery when time-limit risks or control triggers, includes: Tasks that meet the priority threshold or have a deadline shorter than the threshold are assigned to the fast channel and bound to a reserved quota. Tasks within a set range are assigned to the shared channel and scheduled according to their deadlines. Tasks below the lower priority limit are assigned to the low-power channel and configured to be downgraded. When a time-limit risk or control trigger is detected, an interrupt request is issued to elevate the preemption level of the relevant task, the checkpoint is saved, and the quotas of the central processing unit, neural network processing unit, storage bandwidth and bus channel are switched to fast-track tasks. After the triggering conditions are lifted, the preempted task is restored according to the checkpoint, and the channel ownership is re-determined according to the dual gating mechanism.

7. The system according to claim 1, characterized in that, The process of constructing a task energy consumption model based on sensor operating time and processing load to generate an energy consumption budget includes: A parameterized energy consumption model is established, consisting of a sensor activity duration term and a processing load nonlinear term, wherein the nonlinear term is characterized by a quadratic weight of the processing load. The model parameters are calibrated and updated online based on the measurement information provided by the edge computing processing module and the control execution module; Based on the task energy consumption model, task-level and system-level energy consumption budgets are generated, and the energy consumption budgets are provided as constraints to the resource orchestration module and the task modeling module.

8. The system according to claim 7, characterized in that, The step of providing the energy consumption budget as a constraint to the resource orchestration module in order to adjust the sampling frequency, processing frequency, and power domain state includes: The energy consumption budget is mapped to a constraint table of sampling frequency upper limit, processing frequency range and power domain state, and bound to the quotas of central processing unit, neural network processing unit, storage bandwidth and bus channel. According to the constraint table, the low-power channel tasks are downsampled and down-frequencyed, the shared channel is allocated quotas according to the budget weighting, and the fast channel maintains reserved quotas and limits the lower bound of the processing frequency of non-critical segments. The adjustment of sampling frequency, processing frequency and power domain state is stabilized by using a smooth window and hysteresis threshold, and takes effect in the predetermined switching order when the budget is updated.

9. The system according to claim 1, characterized in that, The system also includes: The symbol table storage module is used to store and load model weights, operators, and graph structure metadata in a lightweight symbol table. The symbol table storage module adopts relocatable indexes and variable-length encoding and supports constant pool sharing and differential encoding to provide on-demand dereferencing for lightweight container runtime submodules.

10. A multi-task dynamic scheduling method based on multi-source sensing edge computing according to any one of claims 1 to 9, characterized in that, include: Acquire data from multiple sensors and perform time-base alignment on the acquired data to form a multi-source data stream; The multi-source data stream is preprocessed, modeled, and fused to generate events and features for task generation. Based on the events and characteristics, a task object is generated, and a set of attributes including priority, deadline, energy consumption coefficient, and resource requirement parameters is set. A task energy consumption model is constructed based on sensor operating time and processing load, generating task-level and system-level energy consumption budgets, and mapping the energy consumption budgets to constraints such as upper limit of sampling frequency, processing frequency range and power domain state. Based on the attribute set and energy consumption budget, the central processing unit, neural network processing unit, storage bandwidth and bus channel are jointly orchestrated, and interrupt preemption and checkpoint recovery are performed when time limit risks or control triggers occur. Based on the execution plan generated by the joint orchestration, start / stop and parameter commands are sent to external devices through the industrial communication interface, and the execution status is collected and sent back to update the attribute set and energy consumption budget for subsequent scheduling and control execution.

Images