Industrial internet of things-based device inspection method, device, apparatus and medium

Abstract

This application discloses a method, apparatus, device, and medium for equipment inspection based on the Industrial Internet of Things (IIoT), relating to the field of equipment inspection. The method includes: determining the equipment's state classification result and the corresponding state confidence score for a lightweight model corresponding to any edge node and real-time sensor data of the equipment; when the state confidence score is less than a preset threshold within a preset time period, marking the real-time sensor data corresponding to the state confidence score within that time period as suspicious samples; adding the suspicious samples to the training sample set of an initial fault diagnosis model and updating the initial fault diagnosis model, determining the weight difference before and after the initial fault diagnosis model update, and updating the lightweight model corresponding to the edge node based on the weight difference. This application aims to solve the problem of decreased equipment state inference accuracy caused by the lightweight model's inability to adapt to dynamic environments.

Description

Technical Field

[0001] This application relates to the field of equipment inspection technology, and in particular to a method, device, equipment and medium for equipment inspection based on the Industrial Internet of Things. Background Technology

[0002] In industrial IoT scenarios, equipment inspection typically relies on edge nodes to process real-time sensor data to quickly determine the equipment status. Existing equipment inspection technologies mostly employ large models pre-trained in the cloud, which are then compressed and distributed to edge nodes, enabling edge nodes to perform real-time fault detection with limited computing power.

[0003] However, in real factory environments, external data changes (such as changes in processing materials and seasonal temperature fluctuations). Lightweight models deployed on edge nodes, due to their limited model size and computing power, cannot effectively learn and adapt to these changes, resulting in model drift, which leads to a decrease in model classification accuracy and misjudgment of equipment status. Summary of the Invention

[0004] The main purpose of this application is to provide a device, apparatus, equipment and medium for equipment inspection based on the Industrial Internet of Things, which aims to solve the technical problem of decreased accuracy of equipment state inference caused by the inability of lightweight models to adapt to dynamic environments.

[0005] To achieve the above objectives, this application provides a device inspection method based on the Industrial Internet of Things (IIoT). This method is applied to an IIoT-based device inspection apparatus, which includes a management platform, a sensor network platform, and an object platform connected sequentially. The object platform includes several edge nodes. The management platform is used to train an initial fault diagnosis model using historical sensor data of the device, and generates several lightweight models based on the initial fault diagnosis model and distributes them to the corresponding edge nodes. The edge nodes are used to determine the device status based on the lightweight models and the device's real-time sensor data. The device is an industrial physical entity that requires operational status monitoring and fault diagnosis. The method, executed by the management platform, includes: determining the state classification result of the device and the state confidence score corresponding to the state classification result for the lightweight model and real-time sensor data of the device corresponding to any edge node; when the state confidence score is less than a preset threshold within a preset time period, marking the real-time sensor data corresponding to the state confidence score within the time period as a suspicious sample; adding the suspicious sample to the training sample set of the initial fault diagnosis model and updating the initial fault diagnosis model, determining the weight difference before and after the update of the initial fault diagnosis model, and updating the lightweight model corresponding to the edge node based on the weight difference.

[0006] Optionally, marking the real-time sensor data corresponding to the state confidence level within the time period as suspicious samples includes: acquiring the real-time sensor data corresponding to the time period as a dataset to be screened; calculating the attention level of the real-time sensor data in any dimension of the dataset to be screened, wherein the attention level is used to characterize the temporal changes of the real-time sensor data in the dimension; determining an attention level threshold based on the attention levels of the real-time sensor data in all dimensions of the dataset to be screened, and designating at least one dimension of the real-time sensor data that exceeds the attention level threshold as suspicious samples.

[0007] Optionally, the level of attention is represented as follows:

[0008] in, Indicates the level of attention; The mean of the second-order difference sequence; Indicates deviation; Represents the sum of squared deviations; Indicates degrees of freedom.

[0009] Optionally, updating the lightweight model corresponding to the edge node based on the weight differences includes: extracting a portion of the weight differences corresponding to the lightweight model from the weight differences; performing lightweight processing on the portion of the weight differences to obtain an update increment; and sending the update increment to the edge node so that the edge node adds the update increment to the corresponding lightweight model.

[0010] Optionally, the method further includes: dynamically adjusting the next update time interval of the lightweight model corresponding to the edge node based on the weight difference.

[0011] Optionally, determining the state classification result and state confidence of the device based on the lightweight model corresponding to the edge node and the real-time sensor data of the device includes: inputting the real-time sensor data of the device into the lightweight model to obtain the state classification result corresponding to the sensor in each dimension of the device and the classification probability corresponding to the state classification result; and determining the state confidence of the device based on the classification probability corresponding to the sensor in all dimensions.

[0012] Optionally, generating several lightweight models based on the initial fault diagnosis model includes: performing network pruning and / or quantization compression on the initial fault diagnosis model; and generating several lightweight models based on the network-pruned and / or quantized compressed initial fault diagnosis model.

[0013] Furthermore, to achieve the above objectives, this application also provides an equipment inspection device based on the Industrial Internet of Things (IIoT), comprising: a management platform, a sensor network platform, and an object platform connected in sequence via communication. The object platform includes several edge nodes. The management platform is used to train an initial fault diagnosis model using historical sensor data of the equipment, and to generate several lightweight models based on the initial fault diagnosis model. The edge nodes are used to determine the equipment status based on the lightweight models and real-time sensor data of the equipment. The equipment is an industrial physical entity that requires operational status monitoring and fault diagnosis. The management platform includes: a status confidence acquisition module, used for acquiring confidence scores for any edge node. The system determines the state classification result and state confidence of the device based on the lightweight model corresponding to the edge node and the real-time sensor data of the device; a suspicious sample construction module is used to mark the real-time sensor data corresponding to the state confidence within a preset time period as suspicious samples when the state confidence is less than a preset threshold within a preset time period; a model update module is used to add the suspicious samples to the training sample set of the initial fault diagnosis model and update the initial fault diagnosis model, determine the weight difference before and after the initial fault diagnosis model is updated, and update the lightweight model corresponding to the edge node based on the weight difference.

[0014] This application also provides an equipment inspection device based on the Industrial Internet of Things, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the methods in any of the above possible implementations.

[0015] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method in any of the above possible implementations.

[0016] Compared with the prior art, this application achieves the following technical effects: This application proposes a method, apparatus, equipment, and medium for equipment inspection based on the Industrial Internet of Things (IIoT). First, an initial fault diagnosis model is trained using historical data through a management platform. Then, a lightweight model suitable for edge nodes is generated from this initial model. Next, the state confidence level of the equipment obtained from the lightweight model is evaluated. State confidence level characterizes the certainty of the lightweight model's judgments. When the confidence level remains low for a certain period, the sensor data within that period is marked as suspicious samples. Second, the initial fault diagnosis model is updated using these suspicious samples. Only the weight differences before and after the update are lightweighted and sent back to the corresponding edge nodes to complete the update of the lightweight model. This allows the lightweight model to quickly adapt to changes in external influencing factors in the factory, improving the accuracy of equipment inspection. Attached Figure Description

[0017] Figure 1 This is one of the flowcharts of an industrial Internet of Things-based equipment inspection method according to one embodiment of this application; Figure 2 This is a second flowchart of an industrial Internet of Things-based equipment inspection method according to one embodiment of this application; Figure 3 This is a flowchart of a device inspection method based on the Industrial Internet of Things according to one embodiment of this application; Figure 4 This is a flowchart of a device inspection method based on the Industrial Internet of Things according to one embodiment of this application; Figure 5 This is a structural block diagram of an industrial Internet of Things-based equipment inspection device according to one embodiment of this application; Figure 6 This is a schematic diagram of the structure of an industrial Internet of Things-based equipment inspection device according to one embodiment of this application.

[0018] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0020] In industrial IoT scenarios, equipment inspection typically relies on edge nodes to process real-time sensor data from devices to quickly determine their status. Existing equipment inspection technologies mostly employ large models pre-trained in the cloud, which are then compressed and distributed to edge nodes, enabling these nodes to perform real-time fault detection with limited computing power.

[0021] However, in real factory environments, external data changes (such as changes in processing materials and seasonal temperature fluctuations). Lightweight models deployed on edge nodes, due to their limited model size and computing power, cannot effectively learn and adapt to these changes, resulting in model drift, which leads to a decrease in model classification accuracy and misjudgment of equipment status.

[0022] To address the aforementioned issues, this application provides a method, apparatus, equipment, and medium for equipment inspection based on the Industrial Internet of Things (IIoT). The solution described in this application is detailed below.

[0023] Figure 1 This is one of the flowcharts of an industrial Internet of Things (IIoT) based equipment inspection method according to one embodiment of this application. The equipment inspection method is applied to an IIoT-based equipment inspection device, which includes a management platform, a sensor network platform, and an object platform that are connected in sequence. The object platform includes several edge nodes. The management platform is used to train an initial fault diagnosis model using historical sensor data of the equipment, and to generate several lightweight models based on the initial fault diagnosis model. The edge nodes are used to collect real-time sensor data of the equipment and input the real-time sensor data into the lightweight models to determine the equipment status, i.e., whether the equipment has failed. The equipment inspection method can be executed by the management platform.

[0024] Among them, equipment refers to industrial physical entities in the industrial Internet of Things (IIoT) scenario that require operational status monitoring and fault diagnosis, including but not limited to: industrial robots, CNC machine tools, compressors, fans, pumps, conveying devices, generator sets, etc.

[0025] The initial fault diagnosis model is a large-scale neural network model deployed on the management platform, which can learn the fault characteristic patterns of the equipment from historical sensor data. The lightweight model is a lightweight model generated based on the initial fault diagnosis model through model compression methods. It has low computational complexity and can be deployed on edge nodes with limited computing power.

[0026] It should be noted that the initial fault diagnosis model can be first subjected to lightweight processing, which includes network pruning and / or quantization compression; then, several lightweight models can be generated based on the initial fault diagnosis model after network pruning and / or quantization compression.

[0027] Understandably, network pruning refers to reducing the number of model parameters by removing less important connections or neurons, while quantization compression refers to reducing the storage size of the model by decreasing the numerical precision of the model parameters (e.g., from 32-bit floating-point numbers to 8-bit integers).

[0028] Reference Figure 1 The equipment inspection method based on the Industrial Internet of Things can include the following steps: S11. For any edge node, determine the state classification result of the device and the state confidence corresponding to the state classification result based on the lightweight model and real-time sensor data of the device.

[0029] In this step, the state classification result refers to the lightweight model's judgment on the current operating state of the device. In this embodiment, the state classification result includes, for example, "normal", "warning" and "fault". In other embodiments, it can be reset by the implementer. The state confidence score is used to characterize the credibility of the classification result. The value ranges from 0 to 1. The larger the value, the higher the credibility of the model's judgment.

[0030] In one embodiment, reference is made to... Figure 2 , Figure 2 This is a second flowchart of an industrial IoT-based equipment inspection method according to one embodiment of this application. In step S11, the state classification result of the equipment and the state confidence degree corresponding to the state classification result are determined for the lightweight model corresponding to any edge node and the real-time sensor data of the equipment. This includes the following steps: S111. Input the real-time sensor data of the device into the lightweight model to obtain the state classification results of the sensors in each dimension of the device and the classification probability corresponding to the state classification results. S112. Determine the state confidence of the device based on the classification probability.

[0031] In the specific implementation process, edge nodes continuously collect real-time sensor data from the device (such as temperature, vibration, current, etc.), and input the preprocessed real-time sensor data into a lightweight model deployed locally on the device. This model outputs the state classification results corresponding to the real-time sensor data in each dimension, as well as the classification probabilities corresponding to the state classification results. The category corresponding to the lowest classification probability is taken as the state classification result, and the device's state confidence is further determined based on this classification probability. In other embodiments, the average of all classification probabilities can also be taken as the current state confidence of the device.

[0032] It is understandable that when the classification probability of any dimension is low, it indicates that the model's judgment of the real-time sensor data status of that device in that dimension is uncertain.

[0033] It should be noted that the lightweight model includes a temporal feature extraction layer, a multi-dimensional attention gating module, and a classification head. The temporal feature extraction layer employs a depthwise separable convolutional network structure, responsible for extracting local spatiotemporal features from the original multi-dimensional sensor temporal data. This layer first performs dimensionality reshaping on the input data, converting the original data of shape (time window length L, sensor dimension D) into an image tensor format of (1, L, D). Then, it sequentially passes through two sub-layers: depthwise convolution and pointwise convolution. In the depthwise convolution stage, a convolutional kernel is independently assigned to each sensor dimension. This kernel slides along the time axis, extracting local change patterns of each dimension at adjacent time points, such as abrupt changes in vibration signals or gradient changes in temperature. Because the parameters are not shared between dimensions, the independent temporal characteristics of different physical quantities are preserved. In the pointwise convolution stage, a 1×1 convolutional kernel is used to linearly combine the depthwise convolution outputs of each dimension, achieving the fusion of cross-dimensional information. Depthwise separable convolution reduces the number of parameters in standard convolution from L×D×k. 2 Reduced to D×k 2 +L×D, where k is the kernel size, reducing the number of parameters by approximately 70% to 90% in typical configurations. The output of this layer is a feature map of shape (temporal feature length, sensor dimension). Each feature vector contains both the variation pattern of that sensor dimension within the local time window and incorporates contextual information from other dimensions, providing a high-quality basic feature representation for subsequent attention mechanisms.

[0034] The multi-dimensional attention gating module dynamically recalibrates the features output from the temporal feature extraction layer, enabling the model to adaptively adjust the attention level to each sensor dimension based on the current environmental state. First, global average pooling is performed on the feature map output from the temporal feature extraction layer, compressing all feature values ​​of each sensor dimension over the temporal length into a scalar, resulting in a global description vector of shape D for each sensor dimension. This vector represents the overall activation level of each sensor dimension at the current moment. Subsequently, this vector enters the efficient channel attention submodule. This submodule does not perform dimensionality reduction compression on the vector but directly passes it through a one-dimensional convolutional layer to achieve local cross-channel information interaction. The kernel size k of the one-dimensional convolution is adaptively determined based on the number of channels D; when D ≤ 10, k = 3. After mapping by the Sigmoid activation function, the original attention weights for each channel are output. Building upon this, a difference-aware gating submodule is simultaneously introduced. This submodule maintains a locally updatable distribution baseline vector and calculates the feature distribution difference between the current global description vector and this baseline vector in real time. This difference is then converted into a gating adjustment factor through cosine similarity calculation. This factor is multiplied element-wise with the original attention weights output by the efficient channel attention module, achieving dynamic calibration of the attention weights. Specifically, when the data distribution of a certain dimension significantly deviates from the baseline distribution, the gating factor lowers the final weight of that dimension, thereby suppressing the negative impact of unreliable features on subsequent classification results. Finally, the calibrated attention weights are multiplied channel-wise with the original input features, outputting an adaptively enhanced feature representation. The contribution of each sensor dimension in this representation has been jointly adjusted according to the current environmental state and distribution differences, enabling the model to maintain stable fault detection capabilities in dynamic industrial environments. The classification head is responsible for mapping the enhanced features output by the multi-dimensional attention gating module to specific equipment state classification results and their corresponding classification probabilities. The classification head consists of a fully connected layer and a Softmax activation function. The input dimension of the fully connected layer is equal to the dimension of the augmented features (i.e., sensor dimension D), and the output dimension is equal to the preset number of state categories K (K=3 in this embodiment, corresponding to the three categories "normal", "warning", and "fault"). The fully connected layer linearly combines the augmented features of each sensor dimension into K logical values ​​for each category using a trainable weight matrix and bias vector. Each logical value represents the raw score that the model tends to consider the device to belong to a certain category when judging that sensor dimension alone. Subsequently, the Softmax function normalizes all K logical values, converting them into a probability distribution that satisfies the probability axiom, i.e., the classification probability corresponding to each category, with a value range of (0, 1) and the sum of the probabilities being 1. For a single sensor dimension, the classification head outputs a probability vector of dimension K, where the category corresponding to the maximum probability is the state classification result under that dimension. This maximum probability itself serves as a deterministic measure of the judgment for that dimension.For the device as a whole, the model integrates the classification probabilities output by each of the D sensor dimensions: It adopts a minimum-value strategy, selecting the lowest classification probability among all dimensions as the overall device state confidence score, since uncertainty in any dimension reduces the overall reliability of the judgment; alternatively, in other embodiments, the average of the classification probabilities of all dimensions can be used as the device state confidence score. Ultimately, the overall device state classification result is consistent with the category corresponding to the lowest probability among all dimensions, ensuring the most conservative reliability guarantee for the reported state judgment.

[0035] S12. When the state confidence level is less than the preset threshold within a preset time period, the real-time sensor data corresponding to the state confidence level within the preset time period is marked as a suspicious sample.

[0036] Among them, questionable samples refer to sensor data that cause the lightweight model to make low-confidence judgments. These data may contain new feature patterns that the model has not learned, such as data changes caused by external factors such as changes in processing materials or seasonal temperature changes.

[0037] In the specific implementation process, the status confidence level of the device at each time point is monitored in real time, and the status confidence level within the most recent preset time period is recorded. If the confidence level within this time period is consistently lower than the preset threshold, it is determined that the data in the current stage may have abnormal characteristics that are difficult for the model to identify, and all real-time sensor data within this time period are marked as suspicious samples. In this embodiment, the preset threshold is set to 0.6.

[0038] It should be noted that the status classification results are used to record the current operating status of the equipment. The management platform stores the status classification results in association with the status confidence level to form an inspection log. Subsequently, the frequency of occurrence of each status can be statistically analyzed based on the status classification results, or targeted analysis can be performed on low-confidence samples of specific status categories.

[0039] In one embodiment, reference is made to... Figure 3 , Figure 3 This is the third flowchart of an implementation of an industrial IoT-based equipment inspection method according to this application. Step S12, marking the real-time sensor data corresponding to the state confidence level within a time period as questionable samples, specifically may include: S121. Obtain the real-time sensor data corresponding to the time period as the dataset to be filtered. S122. Calculate the attention level of real-time sensor data for each dimension in the dataset to be screened, so as to characterize the temporal changes of real-time sensor data in that dimension. S123. Determine the attention threshold based on the attention of real-time sensor data in all dimensions of the dataset to be screened, and take real-time sensor data in at least one dimension that exceeds the attention threshold as suspicious samples.

[0040] In the specific implementation process, the multi-dimensional sensor data at all times within the time period when the confidence level is lower than the preset threshold constitutes the dataset to be screened.

[0041] Taking real-time sensor data of any dimension as an example, the real-time sensor data of that dimension within a time period are constructed into a sequence in chronological order. ,in, This represents the number of data points within the time period. Indicates the first The sensor readings at each time step are taken. A second-order difference is calculated from this sequence, where the first-order difference can be expressed as:

[0042] The second-order difference is expressed as:

[0043] The second-order difference sequence is represented as: ,in, , representing the length of the second-order difference sequence; .

[0044] Finally, the variance of the second-order difference sequence is calculated as the attention level of the real-time sensor data in this dimension, as detailed below:

[0045] in, Indicates the level of attention; The mean of the second-order difference sequence; Indicates deviation; Represents the sum of squared deviations; Indicates degrees of freedom.

[0046] Understandably, attention level is used to reflect the severity of data fluctuations; the higher the attention level, the more abnormal the fluctuations in that dimension of data within that time period.

[0047] Furthermore, the average attention level of all dimensions is calculated as the attention level threshold. The attention level of each dimension is compared with this threshold, and all dimensions with attention levels greater than this threshold are selected. The complete sequences of these dimensions within the time period are then used as suspicious samples.

[0048] Understandably, by calculating attention levels, only sensor data suspected of exhibiting abnormal fluctuations can be uploaded, rather than all data, thereby reducing the amount of data uploaded and improving efficiency. This application provides two embodiments: one embodiment directly obtains suspicious samples through state confidence, and the other embodiment calculates attention levels based on state confidence to obtain suspicious samples. In specific implementation, the method for obtaining suspicious samples can be selected by the implementer.

[0049] S13. Add the suspicious samples to the training sample set of the initial fault diagnosis model and update the initial fault diagnosis model. Determine the weight difference before and after the initial fault diagnosis model is updated, and update the lightweight model corresponding to the edge node based on the weight difference.

[0050] Updating the initial fault diagnosis model refers to fine-tuning the trained initial fault diagnosis model using newly added suspicious samples, so that the initial fault diagnosis model can learn newly emerging feature patterns without affecting its original judgment ability.

[0051] Understandably, retraining involves training the model from scratch using all the data, which is computationally intensive, time-consuming, and requires storing all historical data. In contrast, updating adjusts the model parameters using only a small number of new samples, building upon an existing model. This is computationally less demanding, faster, and preserves the model's existing judgment capabilities. This embodiment uses model updating, which allows for faster adaptation to data changes caused by external factors and avoids the resource waste associated with retraining.

[0052] In the specific implementation process, the labeled suspicious samples are added to the training dataset of the initial fault diagnosis model, and the initial fault diagnosis model is fine-tuned several times. After training is completed, the updated model parameters are obtained, and the difference between the model parameters before and after the update is calculated to obtain the weight difference matrix.

[0053] It is understood that the model parameters in this embodiment refer to the model parameters of existing neural networks, including weights and biases.

[0054] In one embodiment, reference is made to... Figure 4 , Figure 4 This is the fourth flowchart of an implementation of an industrial IoT-based equipment inspection method according to this application. In step S13, the lightweight model corresponding to the edge node is updated based on the weight difference, including the following steps: S131. Extract the partial weight difference corresponding to the lightweight model from the weight difference; In this step, after completing the incremental update of the initial fault diagnosis model, the management platform obtains a complete weight difference matrix between the model parameters before and after the update. This matrix records the parameter changes of each network layer in this update. Since the lightweight model deployed on the edge nodes is obtained by pruning and quantizing the initial fault diagnosis model, the network structures of the two are corresponding but not completely identical—the lightweight model retains the key connections and important channels of the original model while removing redundant neurons. Therefore, the management platform needs to extract only the weight differences corresponding to the network layers, channels, and connections retained by the lightweight model from the complete large model weight difference matrix, based on the structure index mapping table of the lightweight model. Specifically, the platform iterates through the parameters of each layer of the lightweight model, finds their position number in the original structure of the large model, and filters out the corresponding difference values ​​from the corresponding large model weight difference tensor according to the index sequence of the preserved channels recorded during pruning. Finally, a partial weight difference matrix with dimensions that perfectly match the lightweight model is obtained. Each element in this matrix corresponds precisely to the amount of change that a certain parameter to be updated in the lightweight model should be adjusted, while the difference values ​​corresponding to the parameters that have been removed during the pruning process are directly discarded.

[0055] S132. Lighten the weight differences to obtain the update increment; In this step, to further reduce the amount of data transmitted to edge nodes and ensure that the update increment is compatible with the numerical accuracy specifications of the lightweight model, the management platform needs to perform the same lightweight processing operation as the initial model compression stage on the extracted weight differences. Specifically, the platform sequentially performs quantization compression and sparse encoding on this part of the weight difference matrix: In the quantization compression stage, the original 32-bit floating-point weight difference values ​​are converted into 8-bit integer representations. The conversion process first calculates the maximum and minimum values ​​in the weight difference matrix, and then linearly maps each difference value to the integer range of 0 to 255. Since the weight differences are usually distributed in a small range near zero, the quantization process can effectively retain their key information while compressing the storage volume to one-quarter of the original. In the sparse encoding stage, the platform performs threshold filtering on the quantized difference matrix, directly setting small difference values ​​with absolute values ​​below a preset threshold to zero, and uses a compressed sparse row format to encode and store the matrix, retaining only non-zero difference values ​​and their position indices. The update increment obtained after the above lightweight processing is usually only 5% to 15% of the original weight difference, which can reduce the network transmission burden. At the same time, it ensures that the numerical precision in the incremental data is completely consistent with the quantization format currently used by the lightweight model on the edge node, so that the edge node can directly perform addition and merging without additional format conversion.

[0056] S133. The update increment is sent to the edge nodes so that the edge nodes can add the update increment to the corresponding lightweight model.

[0057] In this step, after the management platform completes the lightweighting process, it packages the generated update increment with the corresponding edge node identifier and distributes it to the target edge node through the sensor network platform. During the distribution process, the platform adds verification information (such as the hash value of the increment data) to ensure the integrity of the transmission. Simultaneously, it dynamically adjusts the transmission rate based on the network conditions of the edge node to avoid congestion in the industrial IoT network caused by the instantaneous distribution of large amounts of data. Upon receiving the update increment, the edge node first unpacks and sparsely reconstructs the compressed increment data to restore the update tensor, which is completely consistent with the parameter dimensions of the local lightweight model. Then, the node performs element-wise addition of the locally stored lightweight model parameters with the received update increment. That is, for each parameter in the model, its current value is added to the corresponding update increment value to obtain the updated parameter value. This addition operation is performed only within the original parameter space of the lightweight model, without changing the model's structure or adding additional parameter dimensions; therefore, the update time is typically controlled at the millisecond level. After the update is complete, the edge nodes load the new model parameters into the inference engine and continue to process the subsequent real-time sensor data with the updated model. At the same time, they send an update confirmation signal to the management platform, reporting the update timestamp and model version number, so that the management platform can maintain the consistency of the model state of each edge node.

[0058] It should be noted that since the lightweight model is obtained by compressing the initial fault diagnosis model, its parameters correspond to those of the initial fault diagnosis model.

[0059] In the specific implementation process, a portion of the weight differences corresponding to the lightweight model parameters is extracted from the weight difference matrix. This portion of the weight differences is then subjected to the same lightweighting process as the initial fault diagnosis model to generate an update increment, which is then sent to the edge nodes.

[0060] Furthermore, after receiving the update, the edge node adds the update increment to the current parameters of the lightweight model to complete the update of the lightweight model.

[0061] Furthermore, it should be noted that after the update increment is sent to the edge node so that the edge node adds the update increment to the corresponding lightweight model, the method further includes: dynamically adjusting the next update interval of the lightweight model corresponding to the edge node based on the weight difference. Specifically, the management platform first performs global statistics on the weight difference matrix and calculates its L2 norm or absolute mean as the quantification value of the change magnitude of this update. The magnitude of this quantification value directly reflects the degree of influence of the current industrial environment changes on the model parameters. When the weight difference is large, it indicates that the operating conditions of the equipment or environmental conditions have changed significantly (such as the change of processing materials or the change of ambient temperature caused by seasonal changes). The model needs to receive cloud updates more frequently to maintain accurate judgment of the equipment status. When the weight difference is small, it indicates that the current environment is relatively stable and the existing parameters of the lightweight model can still effectively adapt to the real-time data distribution. The update cycle can be appropriately extended to reduce the overhead of network communication and computing resources. Based on the above quantification value, the management platform uses a preset mapping function or PID control strategy to calculate the new update interval: for example, setting a baseline update interval. and adjustment coefficient new time interval ,in, This is a measure of the difference in weights after normalization. As the difference increases, Decrease the interval and increase it; conversely, increase it. To ensure device stability, upper and lower boundary constraints are set for the update interval, such as a minimum interval of 5 minutes and a maximum interval of 24 hours. After adjustment, the management platform records the new interval parameter in the configuration information of the corresponding edge node and uses this new interval as the length of the "preset time period" in subsequent confidence monitoring processes, thus forming a closed-loop adaptive adjustment mechanism that allows the update frequency of the lightweight model to automatically match the drastic changes in the industrial environment.

[0062] In one embodiment, the processor can also dynamically adjust the time interval for triggering the next update based on the magnitude of the difference in weights currently being updated. For example, when the weight difference is large, it indicates that the external influencing factors have changed significantly and the model has been adjusted considerably, so the monitoring time window for the next update is shortened; when the weight difference is small, it indicates that the external influencing factors are relatively stable, so the monitoring time window for the next update is extended.

[0063] This application proposes an equipment inspection method based on the Industrial Internet of Things (IIoT). First, an initial fault diagnosis model is trained using historical data through a management platform. Then, a lightweight model suitable for edge nodes is generated from this initial model. Next, the state confidence level of the equipment obtained from the lightweight model is evaluated. State confidence level characterizes the certainty of the lightweight model's judgments. When the confidence level remains low for a certain period, the sensor data within that period is marked as suspicious samples. Finally, the initial fault diagnosis model is updated using these suspicious samples. Only the weight differences before and after the update are lightweighted and sent back to the corresponding edge nodes to complete the update of the lightweight model. This allows the lightweight model to quickly adapt to changes in external influencing factors in the factory, improving the accuracy of equipment inspection.

[0064] Furthermore, this application embodiment also uses questionable samples to update the initial fault diagnosis model instead of retraining it, which reduces the computational burden on the management platform and improves the inspection efficiency of industrial equipment.

[0065] Based on the above embodiments, Figure 5 This is a structural block diagram of an industrial Internet of Things-based equipment inspection device according to one embodiment of this application, such as... Figure 5 As shown, the equipment inspection device 500 based on the Industrial Internet of Things may include a management platform 501, a sensor network platform 502, and an object platform 503 connected in sequence. The object platform 503 includes several edge nodes 532. Real-time sensor data of each device is collected and acquired by the object platform 503, and then transmitted to the management platform 501 through the sensor network platform 502. The management platform 501 performs equipment inspection based on the real-time sensor data of each target device. The management platform 501 may include: a state confidence acquisition module 510, a suspicious sample construction module 520, and a model update module 530.

[0066] The state confidence acquisition module 510 is used to determine the state classification result of the device and the state confidence corresponding to the state classification result for the lightweight model and the real-time sensor data of the device corresponding to any edge node; the suspicious sample construction module 520 is used to mark the real-time sensor data corresponding to the state confidence within a preset time period as suspicious samples when the state confidence is less than a preset threshold within a preset time period; the model update module 530 is used to add the suspicious samples to the training sample set of the initial fault diagnosis model and update the initial fault diagnosis model, determine the weight difference before and after the initial fault diagnosis model is updated, and update the lightweight model corresponding to the edge node based on the weight difference.

[0067] In an exemplary embodiment, the suspicious sample construction module 520 is further configured to acquire the corresponding real-time sensor data within the time period as the dataset to be screened; calculate the attention level of the real-time sensor data of any dimension in the dataset to be screened, the attention level being used to characterize the temporal changes of the real-time sensor data of the dimension; determine the attention level threshold based on the attention level of the real-time sensor data of all dimensions in the dataset to be screened, and designate the real-time sensor data of at least one dimension that is greater than the attention level threshold as suspicious samples.

[0068] In an exemplary embodiment, the model update module 530 is further configured to extract a portion of the weight difference corresponding to the lightweight model from the weight difference; perform lightweight processing on the portion of the weight difference to obtain the update increment; and send the update increment to the edge node so that the edge node adds the update increment to the corresponding lightweight model.

[0069] In an exemplary embodiment, the model update module 530 is also used to adjust the update time interval of the lightweight model based on the weight difference.

[0070] In an exemplary embodiment, the attention level in the suspicious sample construction module 520 is obtained by performing second-order difference on a sequence of real-time sensor data of that dimension within a time period.

[0071] In an exemplary embodiment, the state confidence acquisition module 510 is further configured to input the real-time sensor data of the device into the lightweight model to obtain the state classification results and the classification probabilities corresponding to the state classification results of the sensors in each dimension of the device; and determine the state confidence of the device based on the classification probabilities corresponding to the sensors in all dimensions.

[0072] Those skilled in the art should understand that the division of the various modules in the embodiments is merely a logical functional division. In actual applications, they can be fully or partially integrated onto one or more actual carriers. These modules can be implemented entirely in software through processing unit calls, entirely in hardware, or a combination of software and hardware. It should be noted that each module in the equipment inspection device based on the Industrial Internet of Things in this embodiment corresponds one-to-one with each step in the equipment inspection method based on the Industrial Internet of Things in the aforementioned embodiments. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned equipment inspection method based on the Industrial Internet of Things, which will not be repeated here.

[0073] Based on the above embodiments, Figure 6 This is a schematic diagram of the structure of an industrial Internet of Things-based equipment inspection device according to one embodiment of this application, as shown below. Figure 6As shown, the device may include a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a device inspection method based on the Industrial Internet of Things (IIoT). The method includes: for any edge node, determining the device's state classification result and state confidence based on the lightweight model corresponding to the edge node and the device's real-time sensor data; when the state confidence is less than a preset threshold within a preset time period, marking the real-time sensor data corresponding to the state confidence within the time period as suspicious samples; adding the suspicious samples to the training sample set of the initial fault diagnosis model and updating the initial fault diagnosis model, determining the weight difference before and after the initial fault diagnosis model update, and updating the lightweight model corresponding to the edge node based on the weight difference.

[0074] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] Based on the above embodiments, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the equipment inspection method based on the Industrial Internet of Things provided by the above methods. The method includes: for any edge node, determining the state classification result and state confidence of the device based on the lightweight model corresponding to the edge node and the real-time sensor data of the device; when the state confidence is less than a preset threshold within a preset time period, marking the real-time sensor data corresponding to the state confidence within the time period as a suspicious sample; adding the suspicious sample to the training sample set of the initial fault diagnosis model and updating the initial fault diagnosis model, determining the weight difference before and after the initial fault diagnosis model update, and updating the lightweight model corresponding to the edge node based on the weight difference.

[0076] Based on the above embodiments, in another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the device inspection method based on the Industrial Internet of Things provided by the above methods. The method includes: for any edge node, determining the device's state classification result and state confidence level based on the lightweight model corresponding to the edge node and the device's real-time sensor data; when the state confidence level is less than a preset threshold within a preset time period, marking the real-time sensor data corresponding to the state confidence level within the time period as a suspicious sample; adding the suspicious sample to the training sample set of the initial fault diagnosis model and updating the initial fault diagnosis model, determining the weight difference before and after the initial fault diagnosis model update, and updating the lightweight model corresponding to the edge node based on the weight difference.

[0077] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A device inspection method based on the Industrial Internet of Things, characterized in that, The method is applied to an equipment inspection device based on the Industrial Internet of Things (IIoT). The device includes a management platform, a sensor network platform, and an object platform connected in sequence. The object platform includes several edge nodes. The management platform is used to train an initial fault diagnosis model using historical sensor data of the equipment, and to generate several lightweight models based on the initial fault diagnosis model and distribute them to the corresponding edge nodes. The edge nodes are used to determine the equipment status based on the lightweight models and the real-time sensor data of the equipment. The equipment is an industrial physical entity that needs to be monitored for operational status and diagnosed for faults. The method is executed by the management platform and includes: For any edge node, the state classification result of the device and the state confidence corresponding to the state classification result are determined based on the lightweight model and real-time sensor data of the device. When the state confidence level is less than a preset threshold within a preset time period, the real-time sensor data corresponding to the state confidence level within the preset time period is marked as a suspicious sample. The suspicious samples are added to the training sample set of the initial fault diagnosis model, and the initial fault diagnosis model is updated. The weight difference before and after the initial fault diagnosis model is updated is determined, and the lightweight model corresponding to the edge node is updated based on the weight difference.

2. The method as described in claim 1, characterized in that, The step of marking the real-time sensor data corresponding to the state confidence level within the preset time period as suspicious samples includes: Obtain the real-time sensor data corresponding to the preset time period as the dataset to be filtered; Calculate the attention level of real-time sensor data for each dimension in the dataset to be filtered; The attention levels and attention thresholds of real-time sensor data across all dimensions in the dataset to be screened are compared, and real-time sensor data with attention levels greater than the attention threshold are identified as suspicious samples.

3. The method as described in claim 2, characterized in that, The level of attention is expressed as follows: in, Indicates the level of attention; The mean of the second-order difference sequence; Indicates deviation; Represents the sum of squared deviations; Indicates degrees of freedom.

4. The method as described in claim 1, characterized in that, The step of updating the lightweight model corresponding to the edge node based on the weight difference includes: Extract the partial weight differences corresponding to the lightweight model from the weight differences; The weight differences are lightened to obtain the update increment; The update increment is sent to the edge node so that the edge node adds the update increment to the corresponding lightweight model.

5. The method as described in claim 4, characterized in that, After the update increment is sent to the edge node so that the edge node adds the update increment to the corresponding lightweight model, the method further includes: The next update interval of the lightweight model corresponding to the edge node is dynamically adjusted based on the weight difference.

6. The method as described in claim 1, characterized in that, The process of determining the state classification result of the device and the state confidence level corresponding to the state classification result based on the lightweight model corresponding to the edge node and the real-time sensor data of the device includes: The real-time sensor data of the device is input into the lightweight model to obtain the state classification results and the classification probabilities corresponding to the state classification results of the sensors of each dimension of the device. The state confidence of the device is determined based on the classification probability.

7. The method as described in claim 1, characterized in that, The generation of several lightweight models based on the initial fault diagnosis model includes: The initial fault diagnosis model is subjected to network pruning and / or quantization compression; Several lightweight models are generated based on the initial fault diagnosis model after network pruning and / or quantization compression.

8. A device for equipment inspection based on the Industrial Internet of Things, characterized in that, The device includes: a management platform, a sensor network platform, and an object platform connected in sequence via communication. The object platform includes several edge nodes. The management platform is used to train an initial fault diagnosis model using historical sensor data of the device, and to generate several lightweight models based on the initial fault diagnosis model. The edge nodes are used to determine the device status based on the lightweight models and real-time sensor data of the device. The device is an industrial physical entity that requires operational status monitoring and fault diagnosis. The management platform includes: The state confidence acquisition module is used to determine the state classification result of the device and the state confidence corresponding to the state classification result for the lightweight model and real-time sensor data of the device corresponding to any edge node. The questionable sample construction module is used to mark the real-time sensor data corresponding to the state confidence level within the preset time period as questionable samples when the state confidence level is less than a preset threshold within the preset time period. The model update module is used to add the questionable sample to the training sample set of the initial fault diagnosis model and update the initial fault diagnosis model, determine the weight difference before and after the initial fault diagnosis model is updated, and update the lightweight model corresponding to the edge node based on the weight difference.

9. A device inspection device based on the Industrial Internet of Things, characterized in that, The equipment inspection equipment based on the Industrial Internet of Things includes: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images