Equipment fault early warning analysis method, equipment and medium

Abstract

The invention provides an equipment fault early warning analysis method, equipment and a medium, which are applied to a system comprising a remote control end and local monitoring equipment. The method comprises the following steps: a remote control terminal performs feature screening based on an early warning model trained for target equipment to obtain a target feature subset meeting resource constraints of local monitoring equipment; generating a lightweight early warning model according to the target feature subset and issuing the lightweight early warning model to local monitoring equipment; the local monitoring device collects monitoring data corresponding to the target feature subset, analyzes the monitoring data by using a lightweight early warning model and generates fault early warning information; and the remote control terminal performs collaborative updating on the target feature subset and the lightweight early warning model according to the continuously acquired target equipment data. According to the method, high-precision, self-adaptive and sustainable fault early warning on the edge side with limited resources is realized.

Description

Technical Field

[0001] This invention relates to the fields of industrial Internet of Things and predictive maintenance technology, and in particular to a method, device and medium for equipment fault early warning analysis. Background Technology

[0002] With the increasing demand for intelligent operation and maintenance of industrial equipment, fault early warning technology is gradually evolving from centralized back-end analysis to edge deployment. For critical individual equipment such as machine tools and compressors, dedicated embedded fault monitors are often equipped to achieve real-time, local condition assessment and early warning. However, such monitoring devices are usually limited by size, cost, and power consumption, and their computing power, storage space, and energy supply are relatively limited, making it impossible to directly run complex analysis models or process full, high-frequency sensor data.

[0003] Existing technical solutions face multiple challenges: First, training samples for specific devices (such as a machine tool 280) are often scarce or incomplete, resulting in insufficient accuracy and poor generalization ability of the local model trained specifically for it. Second, even after supplementing samples with data augmentation techniques, the model still needs to process high-dimensional feature data during inference, which contradicts the limited processing capabilities of edge devices. Simply collecting and calculating all features will lead to response latency, a surge in power consumption, and even device overload. In addition, traditional feature selection methods mostly rely on offline analysis by backend servers, which cannot dynamically adapt to the real-time constraints and data acquisition conditions of edge devices (such as sensor frequency and acquisition costs), and the selected feature subset may not meet the resource constraints of the edge.

[0004] Therefore, how to build a fault early warning model that is both highly accurate and low-resource-consumption and can be dynamically optimized on resource-constrained local monitoring equipment has become an urgent technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows: According to a first aspect of the present invention, a method for equipment fault early warning analysis is provided, applied to a system comprising a remote control terminal and at least one local monitoring device, wherein each local monitoring device is connected to at least one target device, the method comprising the following steps: S100: The remote control terminal performs feature filtering based on the initial early warning model trained for the target device to obtain a subset of target features.

[0006] S200, the remote control terminal generates a corresponding lightweight early warning model based on the target feature subset and sends it to the corresponding local monitoring device.

[0007] S300, the local monitoring device acquires the monitoring data of the target device corresponding to the target feature subset, processes it using the lightweight early warning model, and generates and outputs fault early warning information for the target device.

[0008] According to a second aspect of the present invention, an electronic device is provided, including a processor and a memory; the processor executes the steps of the method described in the first aspect of the present invention by invoking a program or instructions stored in the memory.

[0009] According to a third aspect of the present invention, a computer-readable storage medium is provided that stores a program or instructions that cause a computer to perform the steps of the method described in the first aspect of the present invention.

[0010] The present invention has at least the following beneficial effects: This invention utilizes a cloud-edge collaborative architecture of remote feature selection and local model deployment. It transforms a computationally complex early warning model that relies on all features into a lightweight model that relies only on key features, simplifying computation. This enables high-performance early warning capabilities to be achieved on local monitoring devices with strictly limited computing, storage, and energy consumption, overcoming the resource bottleneck of deploying early warning models at the edge. Furthermore, by performing feature selection based on local resource constraints through remote control, it fundamentally ensures that the final lightweight early warning model matches the actual processing capabilities of the device. This avoids the risk of model failure or performance degradation at the edge due to insufficient resources, as is common in traditional methods, achieving system-level optimization of early warning performance and resource consumption from the source.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A schematic diagram illustrating an application scenario of an equipment fault early warning analysis method provided in an embodiment of the present invention; Figure 2 A flowchart of a device fault early warning analysis method provided in an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0016] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0017] This invention provides a method for equipment fault early warning analysis, applicable to a system comprising a remote control terminal and at least one local monitoring device. The system architecture can be found in [reference needed]. Figure 1 As shown, it mainly includes: a remote control terminal 1 (such as a cloud server or a background central processing unit), one or more local monitoring devices 2, and target devices 3 (such as specific machine tools, compressors, generators and other industrial equipment) connected to each local monitoring device.

[0018] The local monitoring device 2 is a resource-constrained embedded device, whose typical structure may include a base 201 and indicator lights 202 mounted on it. The base 201 integrates a microprocessor, memory, a data acquisition module (such as an interface supporting industrial protocols like Modbus and OPC UA), a wireless communication module (such as a 4G / 5G or Wi-Fi module), and indicator light driver circuitry. External operation buttons can be used for device reset, manual testing, or mode switching. The indicator lights 202 typically use multi-color LEDs to visually display the real-time status level of the target device (e.g., green for normal, yellow for warning, and red for fault). It is understood that the specific implementation of the local monitoring device can be adjusted according to the application scenario, but its core lies in having limited computing, storage, and communication resources specifically designed to execute lightweight fault warning logic.

[0019] The following combination Figure 2The flowchart shown illustrates a complete embodiment of this method.

[0020] like Figure 2 As shown, the equipment fault early warning analysis method provided in this embodiment of the invention may include the following steps: S100: The remote control terminal performs feature filtering based on the initial early warning model trained for the target device to obtain a subset of target features.

[0021] The remote control terminal first acquires historical operating data of the target device and data of similar devices, expands the samples through data augmentation techniques (such as SMOTE and generative adversarial networks), and trains a high-precision initial early warning model (e.g., a classification model based on XGBoost or deep neural networks).

[0022] Subsequently, feature filtering is performed to adapt to the resource constraints of local monitoring equipment. Specifically, this includes: S101, divide all sensor features of the initial early warning model into multiple logically independent or related feature groups; wherein, the feature group includes at least a seed feature group defined based on prior knowledge.

[0023] All sensor features include, but are not limited to, physical quantities such as temperature, pressure, vibration amplitude, spectral composition, current, voltage, and harmonic distortion rate collected from the target equipment and its process system.

[0024] The division of the feature groups is based on one or more of the following principles: Physical attribute principle: Group sensor characteristics that measure the same physical quantity or are located in the same functional component into one group. For example, divide the temperature measuring points distributed at various points on the bearing housing into one characteristic group; or divide the vibration spectrum characteristics (such as the amplitude of the first harmonic and second harmonic) of the motor drive end into one characteristic group.

[0025] Signal source principle: Group features from the same data acquisition module, communication bus, or with the same sampling frequency together.

[0026] Domain knowledge principle: Based on the prior knowledge of equipment mechanism and fault diagnosis, group together the known features that work together physically or jointly characterize a specific fault mode.

[0027] Among the multiple feature groups that have been divided, at least one seed feature group is defined. The seed feature group is a set of features that are pre-determined based on domain expert knowledge, historical fault data analysis, or physical failure models, and are necessarily strongly correlated with the key failure modes of the target equipment. For example, for bearing fatigue failures in rotating machinery, vibration acceleration characteristics at a specific high-frequency band (such as the bearing's passing frequency) are defined as a seed feature group; for motor insulation aging failures, current harmonic components of a specific order can be defined as a seed feature group. The remaining feature groups are non-seed feature groups.

[0028] S102, for the multiple feature groups, the importance measure of each feature group to the performance of the initial warning model is determined by evaluating the decrease in accuracy of the initial warning model output when different feature groups are masked in the input of the initial warning model.

[0029] In this invention, the precision reduction value for each feature group is determined through the following steps: S1021, In the input data of the initial warning model, all feature data belonging to the feature group are masked (e.g., zeroed or deleted), and the masked data is input into the initial warning model to obtain the corresponding evaluation accuracy such as classification accuracy, F1-Score, etc.

[0030] S1022, calculate the difference between the evaluation accuracy and the baseline accuracy of the initial early warning model when all feature data are input. This difference is the accuracy decrease value corresponding to the feature group. The larger the accuracy decrease value, the more important the feature group is to maintaining the overall accuracy of the initial early warning model.

[0031] In a basic embodiment, for each feature group G, the accuracy reduction ΔAccuracy(G) of the model after masking the data of that group in the initial warning model input is evaluated as its basic importance metric.

[0032] In a preferred embodiment, in order to more comprehensively evaluate the essential contribution of feature groups to fault diagnosis, an evaluation index that integrates multiple dimensions is used to calculate the importance measure.

[0033] For a feature set G, the importance metric I(G) of the feature set G is calculated by the following formula: I(G)=β·△Accuracy(G)+(1-β)·△Separability(G).

[0034] Wherein, ΔAccuracy(G) represents the decrease in accuracy of the initial warning model caused by the masking feature group G.

[0035] ΔSeparability(G) represents the degree of decrease in the distinguishability of the initial warning model for different fault modes caused by the masked feature group G. Here, fault mode refers to a predefined category of equipment state that needs to be distinguished based on domain knowledge, historical fault records, or equipment mechanism. For example, for a machine tool spindle bearing, its fault modes can be predefined as: ["Normal", "Inner Ring Failure", "Outer Ring Failure", "Rolling Element Failure"]. Each training or validation sample has a corresponding fault mode label.

[0036] The specific calculation method for ΔSeparability(G) may include: Step 1: Extract feature representations under the baseline state.

[0037] The initial early warning model is input using a complete failure sample validation set (containing all features). For each sample in the sample set, its corresponding intermediate layer feature representation is extracted. Based on the failure mode label of the sample, these intermediate layer feature representations are grouped to obtain grouped feature vectors.

[0038] The intermediate layer feature representation refers to the feature vector extracted from a specific intermediate layer (such as a fully connected layer or a pooling layer) after the sample data is input into the initial warning model (usually a deep learning model). This vector is regarded as the model's internal representation of the sample.

[0039] Step 2: Calculate the fault mode separability metric S under baseline conditions. full .

[0040] Based on the grouped feature vectors obtained in step 1, an index is calculated that can quantify the degree of separation between different failure modes. Below are examples of two common and implementable calculation methods: Method 1: Separation based on inter-class center distance (1) For each predefined fault mode c (e.g., "normal", "bearing inner ring fault", "rotor imbalance", etc.), calculate the mean of the feature vectors of all samples belonging to the fault mode in the middle layer of the model, and obtain the feature center μ of the fault mode. c .

[0041] (2) Calculate the Euclidean distance between each pair of all failure modes (usually using Euclidean distance, i.e. L2 norm), and obtain the average value of these distances.

[0042] (3) Use this average distance as the fault mode separability measure S of the model in this state. full In one illustrative embodiment, Where K is the total number of predefined fault modes, μ i and μ jLet i and j represent the feature centers of the i-th and j-th fault modes, respectively. The value of i is from 1 to K-1, and the value of j is from i+1 to K. |||| represents the L2 norm of the vector.

[0043] Method 2: Separation based on Fisher's discriminant ratio This method quantifies the model's overall ability to distinguish different failure modes by calculating the ratio of inter-mode divergence to intra-mode divergence. The specific steps are as follows: (1) Calculate the global and individual fault mode feature centers: Calculate the mean of all sample feature vectors to obtain the global feature center μ. global ; For each fault mode c, calculate the mean of the feature vectors of the samples belonging to that fault mode, and obtain the feature center μ of that mode. c .

[0044] (2) Calculate the inter-class scatter matrix Sb, which measures the degree of dispersion between the feature centers of different failure modes. , where n c is the number of samples in category c, and T represents the transpose of a matrix or vector.

[0045] (3) Calculate the intraclass scatter matrix Sw, which measures the degree of clustering (dispersion) of sample features within the same failure mode. , where Dc represents the set of samples belonging to fault mode c, and x is the feature vector of a sample in Dc.

[0046] (4) Based on Sw and Sb, calculate the Fisher discriminant ratio as a measure of separability S. full .

[0047] In one illustrative embodiment, S full For matrix Sw -1 The trace of Sb is the sum of the elements along the main diagonal of the matrix, and Sw is the trace of Sb. -1 This represents the inverse matrix of Sw. The larger this value, the higher the separation between failure modes and the better the clustering within modes.

[0048] Step 3: Extract the feature representation after masking feature group G.

[0049] In the input data of the initial early warning model, all feature data belonging to feature group G are masked (e.g., set to zero). Keeping the model structure and parameters unchanged, the same fault sample validation set is input into the initial early warning model, and the feature representations of the same intermediate layers are extracted. Similarly, grouping is performed according to the fault mode label.

[0050] Step 4: Calculate the fault mode separability metric S after masking feature group G. masked(G).

[0051] Using the exact same method and formula as in step 2, and based on the grouped feature vectors obtained in step 3, calculate the separability measure S after masking. masked (G).

[0052] Step 5: Calculate the failure mode distinguishability degradation rate ΔSeparability(G).

[0053] In an illustrative embodiment, ΔSeparability(G) = (S full -S masked (G)) / S full ΔSeparability(G) is a ratio between 0 and 1 (it may be greater than 1, but it is usually positive). The larger the ΔSeparability(G), the more severe the decrease in the model's internal ability to distinguish different failure modes after masking feature group G, thus implying that feature group is more important for maintaining the model's core diagnostic and discriminative capabilities.

[0054] β is a preset fusion weight parameter used to balance the considerations of accuracy and distinguishability, and its value ranges from 0 to 1. In a specific configuration example, β can be 0.7.

[0055] Under this definition, the value of I(G) comprehensively reflects the importance of feature group G in maintaining the overall performance (accuracy) of the model and its core diagnostic capability (fault mode discrimination). The higher the value of I(G), the more important the feature group G is.

[0056] S103, based on the importance metric, and on the premise of meeting the resource constraints of the local monitoring equipment, select one or more feature groups that make the accuracy decrease of the initial early warning model within a preset cumulative accuracy decrease threshold to form the target feature subset.

[0057] In this invention, resource constraints refer to the objectively existing and insurmountable physical or system limits of the local monitoring equipment. It is an upper limit value in one or more dimensions. A subset of candidate features must satisfy all these constraints; otherwise, it will be directly excluded. For example: Processing capacity constraint: The maximum number of features that the device's CPU can process in a single inference run (e.g., ≤30).

[0058] Memory constraint: The maximum memory (e.g., ≤256MB) that the device can use to store feature data and run the model.

[0059] Communication bandwidth constraint: The maximum size of a data packet that a device can receive in a single transaction.

[0060] Power consumption constraint: The maximum allowable energy consumption of a single sampling-calculation cycle of the device.

[0061] The preset cumulative accuracy decline threshold refers to the maximum allowable performance loss of the model's early warning capabilities during the feature selection process, in order to improve the efficiency of edge-side resources. This threshold is a preset system parameter, and its technical meaning is that the performance (such as classification accuracy) of the lightweight early warning model built from the final selected subset of target features must not exceed this threshold compared to the performance decline of the initial early warning model using all features.

[0062] In one illustrative embodiment, the preset cumulative accuracy decline threshold is set to a fixed absolute value. For example, it can be set to a cumulative accuracy decline of no more than 5%. This means that if the baseline accuracy of the initial warning model on the validation set is 95%, then the accuracy of the lightweight warning model should not be lower than 90%.

[0063] In a preferred embodiment, the threshold is dynamically determined based on the baseline accuracy of the initial early warning model. The design principle is that the stronger the model's baseline discrimination capability (the higher the baseline accuracy), the greater the allowable absolute accuracy loss space in exchange for resource efficiency.

[0064] Specifically, the preset cumulative accuracy decrease threshold Δmax can be dynamically calculated using the following formula: Δmax = f(Acc) baseline ), where Acc baseline Let f() be the baseline accuracy (e.g., 95%) of the initial early warning model on the validation set, and let f() be a preset mapping function. For example, a linear function can be used: Δmax = h × Acc baseline Where h is a scaling factor less than 1 (e.g., 0.05). Accordingly, if the reference accuracy is 95%, the allowable cumulative accuracy decrease threshold is 4.75%.

[0065] This design ensures that for high-performance models, greater resource savings can be achieved while maintaining high absolute accuracy; while for models with average performance, a more conservative strategy is adopted to avoid reducing the reliability of early warnings to an unacceptable level due to excessive feature simplification.

[0066] To achieve a performance loss tolerance preset while meeting resource constraints, this invention proposes a specific screening strategy, including differentiated processing standards for different feature groups: Differentiated judgment threshold: To reflect prior knowledge in the domain, different tolerance thresholds for accuracy degradation are set for seed feature groups and non-seed feature groups, i.e., accuracy degradation judgment thresholds.

[0067] Seed feature group: Due to its strong correlation with the core failure mechanism, a stricter (lower) judgment threshold (e.g., 1%) is set. Once the accuracy drop caused by removing it alone approaches or exceeds this threshold, it is given the highest priority and retained.

[0068] Non-seed feature group: Set a relatively lenient (higher numerical) decision threshold (e.g., 5%). In the decision-making process, its retention priority is lower than that of the seed feature group.

[0069] This strategy ensures that, under the same conditions of decreased accuracy, seed feature groups will receive a higher retention priority, thereby solidifying domain knowledge into the selection logic.

[0070] In a basic embodiment, step S103 is implemented through a step-by-step rule process that explicitly applies a differential judgment threshold. Its core is to use the threshold to determine the removability of each feature group. The specific process is as follows: S1031, Hard Resource Constraint Filtering: Traverse all possible feature group combinations and directly exclude any combinations that violate local monitoring device resource constraints (such as exceeding the total number of features or memory usage limits). This step ensures that all candidate subsets are physically feasible.

[0071] S1032, Threshold-based feature group selection: In the candidate subset that has passed resource filtering, each feature group is evaluated and decided one by one according to the following rules, which directly depends on the judgment threshold preset for its type: 1. Seed Feature Group Evaluation: For each seed feature group G s Obtain its individual importance metric I(G) s If I(G) s If the accuracy decrease value of the seed feature group is greater than the threshold Ts (e.g., 1%), then the group is forcibly retained; otherwise, it can be removed in theory, but since it is a seed feature, it is still given high priority by default.

[0072] Because the threshold set for seed groups is extremely low (e.g., 1%), the importance metrics of the vast majority of seed feature groups will exceed this threshold, thus triggering the mandatory retention rule. This directly and quantitatively achieves their highest retention priority.

[0073] 2. Sorting and filtering of non-seed feature groups: Step 1: Eligibility Screening. For each non-seed feature group G n Only when its importance measure I (G n Features are only considered when their percentage is less than or equal to the non-seed feature group decision threshold Tn (e.g., Tn=5%). This directly applies the threshold to exclude irrelevant or redundant features whose removal alone would cause excessive performance loss.

[0074] Step 2: Prioritization. Sort all eligible non-seed feature groups according to their importance metric from high to low, forming a candidate queue.

[0075] Step 3: Optimal Addition and Global Constraint Check. Following the sorting order, attempt to add non-seed feature groups to the candidate subset that already contains all mandatory seed features. Before adding each feature group, two checks must be performed: a. Cumulative precision decline check: After being added to this group, the cumulative precision decline value of the candidate subset (the sum of the importance metrics of all removed groups) must not exceed the preset cumulative precision decline threshold (e.g., 5%).

[0076] b. Resource constraint review: After joining the group, the candidate subset must still meet all resource constraints.

[0077] If the check passes, add the group; if any check is violated, skip the group and continue evaluating the next one in the queue.

[0078] After the above process, the target feature subset F is obtained. target It will necessarily include all seed feature groups that are forcibly retained because they exceed their low threshold, as well as a batch of non-seed feature groups that are of high importance and can keep the overall performance within the global tolerance.

[0079] In a preferred embodiment, step S103 achieves intelligent search by constructing and maximizing a joint optimization objective function Q. This function satisfies objectives such as performance loss tolerance presets and reflecting high priority of seed features, and is fused and encoded into a unified mathematical problem.

[0080] To ensure that the solution to the joint optimization objective function Q(F) conforms to domain prior knowledge (i.e., prioritizing the retention of seed feature groups), this invention implements a priority fusion strategy in the optimization model. In one embodiment, by introducing a high-weighted missing feature penalty term for each seed feature group in the comprehensive cost term C(F), the cost of removing a seed feature group far outweighs the resource savings, thereby guiding the optimization algorithm to automatically retain it. In another embodiment, by adding hard constraints to the optimization problem, it is mandated that the final feature subset must contain all seed feature groups, integrating domain knowledge as an inviolable rule into the solution process. These two embodiments are described below.

[0081] (Example 1) In this embodiment, the joint optimization objective function Q(F) is defined as: Q(F) = α·A(F) - (1-α)·C(F).

[0082] Where A(F) represents the accuracy retention rate of the initial early warning model when using a subset of candidate features F, and C(F) represents the normalized comprehensive overhead of collecting and processing F on the local monitoring device. α is a trade-off coefficient dynamically adjusted according to the real-time resource status (such as remaining power and CPU load) of the local monitoring device, α∈[0,1]. When device resources are abundant, α approaches 1, and optimization focuses more on accuracy; when resources are scarce, α approaches 0, and optimization focuses more on reducing overhead.

[0083] A resource-sufficient status refers to a situation where the real-time available resources of the local monitoring equipment are significantly higher than the minimum requirements for safe operation, allowing for the allocation of more resources to improve early warning performance. This can be determined through one or more of the following conditions: Ample computing resources: The current utilization of CPU / GPU remains below a low threshold (e.g., <40%).

[0084] Ample memory resources: The current percentage of free memory to total memory is higher than a high threshold (e.g., >50%).

[0085] Ample energy resources: The device's battery charge (or remaining available energy) is above a high threshold (e.g., >70%).

[0086] Ample communication resources: Network bandwidth utilization remains below a low threshold.

[0087] In this state, the strategy may tend to trade resources for performance, so the tradeoff coefficient α is adjusted to be close to 1 (e.g., α=0.9), so that the optimization objective function Q(F) focuses more on maximizing the accuracy retention rate A(F).

[0088] Resource scarcity refers to a situation where the real-time available resources of the local monitoring equipment are approaching their constraint limits. The system must prioritize basic functions and battery life, and must minimize the overhead of early warning tasks. This can be determined through one or more of the following conditions: Computing resources are strained: The current utilization of CPU / GPU is consistently above a high threshold (e.g., >80%).

[0089] Memory resource shortage: The current percentage of free memory is below a low threshold (e.g., <20%).

[0090] Energy shortage: The device's battery charge (or remaining available energy) is below a low threshold (e.g., <20%), or the device is in energy-saving mode.

[0091] Communication resources are scarce: network connections are unstable or bandwidth costs are high.

[0092] In this state, the strategy must sacrifice performance to ensure operation, so the tradeoff coefficient α is adjusted to be close to 0 (e.g., α=0.1), so that the optimization objective function Q(F) focuses more on minimizing the overall cost C(F).

[0093] In this invention, A(F) is used to quantify the degree to which the model performance is preserved relative to the use of all features when using a subset of features F.

[0094] In an illustrative embodiment, A(F) = Acc(F) / Acc(Full). Here, Acc(Full) is the evaluation accuracy (e.g., classification accuracy, F1 score, etc.) of the initial warning model on the validation set when using all features. Acc(F) is the evaluation accuracy of the evaluation model obtained after structural adaptation of the initial warning model based on a subset of candidate features F, on the same validation set. This evaluation model is used only for background performance simulation, and its input layer is adjusted to receive only features from F. A(F) ∈ [0, +∞), typically A(F) ≤ 1.

[0095] In another preferred embodiment, in order to quickly estimate A(F) without retraining the model, A(F) is calculated based on the importance measures of each feature group not included in F, as determined in S102: .

[0096] C(F) is used to quantify the cost required to operate on resource-constrained local monitoring equipment. Its calculation aims to balance two objectives: minimizing conventional resource consumption and ensuring that key features (seed feature sets) required for domain knowledge are preferentially retained through a penalty mechanism.

[0097] In this embodiment, to incorporate domain knowledge that prioritizes the retention of seed feature groups into the automated optimization process, the definition of the normalized synthesis cost C(F) is expanded to include an explicit priority penalty mechanism. The specific definition is as follows: Where M is the number of expense items, ω k E represents the preset weight coefficient for the k-th expense item, and the sum of the preset weight coefficients for all expense items is 1. k (F) The normalized value of the k-th expense item. P seed (F) is a seed feature missing penalty term, used to penalize seed feature groups that are not included in the candidate feature subset F; λ is a preset large positive weight coefficient, used to significantly amplify the influence of the penalty term, thereby ensuring that the algorithm strongly tends to retain seed feature groups during the optimization process.

[0098] In one embodiment of the present invention, the overhead items mainly include, but are not limited to, the following dimensions: Computational resource overhead Ecomp(F): Represents the normalized value of the computational cost required to process the feature subset F. Its value is obtained by dividing the estimated required processor computation (e.g., CPU cycles or floating-point operations) by the upper limit of available computing resources of the local monitoring device (e.g., maximum processor capacity per second).

[0099] Memory usage overhead term Emem(F): Represents the normalized value of memory usage required to store feature data and run the early warning model. Its value is obtained by dividing the estimated peak memory usage (such as RAM usage) by the upper limit of available memory resources of the local monitoring device (such as total available RAM capacity).

[0100] Communication data overhead term Ecom(F): Represents the normalized value of the communication load generated by collecting or receiving feature data. Its value is obtained by dividing the estimated total amount of data to be transmitted in a single cycle by the upper limit of the single-cycle communication capacity of the local monitoring device's communication interface (such as the product of bandwidth and sampling period).

[0101] The timing mismatch penalty overhead term P(F) is a normalized value used to quantify the additional system overhead caused by the inconsistent acquisition cycles of features within the feature subset F. This overhead is not a direct computation or communication cost, but rather a loss of system efficiency due to asynchronous data stream timing. It is mainly reflected in: processing delays caused by waiting for slow feature data, increased scheduling complexity for synchronizing data from different frequencies, and buffer memory overhead. By incorporating this penalty term into the overall overhead C(F), the optimization algorithm can be guided to automatically favor feature combinations with more coordinated acquisition timings, thereby improving the overall response efficiency and operational smoothness of the edge warning system.

[0102] In one embodiment, the estimated actual costs of Ecomp(F), Emem(F), and Ecom(F) can be obtained through theoretical analysis, model profiling, or offline performance simulation. The upper limit of the available resources is the known hardware specifications of the local monitoring device or a preset system quota.

[0103] In one illustrative embodiment, P(F) is calculated as follows: Where Tu and Tv are the inherent acquisition periods (in seconds) of features u and v on the target device, respectively. ρ uv ρ is the historical correlation coefficient (absolute value) between feature u and feature v, calculated by analyzing historical data. uv ∈[0,1]. The higher the correlation (ρ) uv For feature pairs (where the size of the feature pair is larger), if the difference in collection periods is large, the negative impact on data fusion will be greater, and the penalty will be more severe. u,v∈F,u＜v This indicates that the summation is performed on all unordered feature pairs (u, v) in the candidate feature subset F (i.e., each pair is calculated only once).

[0104] In this invention, P seed One specific way to calculate (F) is as follows: Among them, G s Represents a seed feature set, γG s G represents s The corresponding preset penalty coefficient is usually much larger than the normalized value of the regular resource overhead. For example, it can be set to 1 or 10, depending on the range of normalized resource overhead.

[0105] In this invention, the setting of the preset penalty coefficient must follow a core principle: in the decision of the joint optimization function Q(F), the penalty cost of removing any seed feature group must be much greater than the conventional resource overhead that may be saved by retaining the group.

[0106] In a specific configuration example, the weighted sum of regular resource overhead is typically evaluated to be no more than 5. To ensure evidentiary value, each seed feature group G is... s The preset penalty coefficient γG s The value is uniformly set to 100. This means that in the optimization calculation, removing a seed feature group will cause the overall overhead C(F) to increase by at least 100 units, a penalty that is far greater than any possible savings in regular resources (≤5).

[0107] In this invention, the setting of the preset penalty coefficient has the same technical intent as setting an extremely low judgment threshold (such as 1%) for the seed feature group in the rule-based filtering embodiment. Both are aimed at solidifying the domain knowledge with the highest retention priority in the seed feature group into decision logic. Within the optimization framework, a high penalty coefficient is a mathematical means to achieve this priority.

[0108] Therefore, during the search optimization process, if the algorithm considers removing a seed feature group, even if this significantly reduces other resource overhead, it will introduce a large P-value. seed The increase in overall cost C(F) due to the removal of seed feature sets leads to a significant increase in overall cost, which in turn reduces the objective function value Q(F). This forces the optimization algorithm to prioritize retaining seed feature sets unless removing them would bring extreme resource gains (e.g., making a solution that was originally infeasible due to resource constraints feasible). Finally, the remote control terminal uses an optimization algorithm (such as a genetic algorithm) to search for the candidate feature subset that maximizes the value of Q(F) and determines it as the target feature subset.

[0109] (Example 2) In this embodiment, the priority of preserving seed feature groups is achieved by adding hard constraints to the optimization problem without modifying the definition of the overhead term in the objective function. Therefore, the normalized synthesis overhead C(F) in the joint optimization objective function Q(F) = α·A(F)-(1-α)·C(F) contains only the conventional resource overhead term, which is defined as: .

[0110] Meanwhile, in solving the optimization problem of maxQ(F), the following hard constraints are added: Constraints: G s ∈F, ∀G s ∈S.

[0111] Here, S represents the set of all seed feature groups. This constraint mandates that any feasible candidate feature subset F must contain all seed feature groups.

[0112] The key difference between this embodiment and Embodiment 1 (using a penalty term) is that it transforms domain knowledge into an absolute constraint on the solution space. During the search, the optimization algorithm only considers candidate solutions that already include all seed feature groups, and optimizes the selection of non-seed features to maximize Q(F). This method avoids designing complex penalty terms in C(F), instead using a simple and explicit rule to guarantee the preservation of core features with the highest determinism and robustness.

[0113] Both of the above embodiments prioritize the retention of domain knowledge from the seed feature group, successfully integrating it into the solution framework of the joint optimization objective function Q(F). This is achieved by adding a penalty term to the objective function and by adding constraints during the solution process, respectively, so that the optimization results automatically conform to the prior knowledge of the domain.

[0114] S200, the remote control terminal generates a corresponding lightweight early warning model based on the target feature subset and sends it to the corresponding local monitoring device.

[0115] After determining the target feature subset, the remote control terminal performs the following operations to complete model deployment: S201, Lightweight Early Warning Model Generation: Based on the initial early warning model, the remote control terminal adapts and optimizes the structure of the initial early warning model according to the target feature subset to generate a dedicated lightweight early warning model. This adaptation process includes, but is not limited to: adjusting the input layer structure of the initial early warning model according to the target feature subset, and optionally retraining or fine-tuning the initial early warning model to ensure its early warning performance after using simplified features.

[0116] S202, Configuration Information Encapsulation: Simultaneously, the remote control terminal generates configuration information related to the target feature subset and the lightweight early warning model. This configuration information includes at least: a feature list, the acquisition frequency of each feature, data preprocessing parameters (such as normalization coefficients), and hyperparameters required for model inference.

[0117] S203, Model and Configuration Distribution: Finally, the remote control terminal encapsulates the lightweight early warning model and the configuration information, and distributes them together to the local monitoring device corresponding to the target device. Upon receiving the data, the local monitoring device initializes the data acquisition process according to the configuration information and loads the lightweight early warning model to enter a ready state.

[0118] A dedicated, lightweight early warning model is obtained through retraining or adjustment. This model, along with the configuration information of the target feature subset (including feature list, acquisition frequency, and preprocessing parameters), is sent to the corresponding local monitoring equipment.

[0119] S300, the local monitoring device acquires the monitoring data of the target device corresponding to the target feature subset, processes it using the lightweight early warning model, and generates and outputs fault early warning information for the target device.

[0120] The fault warning information can be used to drive the human-machine interaction device, for example, to control the indicator lights set on the local monitoring device to change their display status, so as to provide intuitive fault level prompts.

[0121] Specifically, the process of using the lightweight early warning model to generate and output fault early warning information for the target device includes: S301, the lightweight early warning model calculates a preliminary failure probability based on the input monitoring data.

[0122] The local monitoring device inputs the acquired monitoring data (organized and preprocessed according to a subset of target features) into the lightweight early warning model. This model performs forward computation and outputs a numerical value representing the probability of a fault occurring in the current state, called the preliminary fault probability P. pre The value range is usually [0, 1].

[0123] S302, based on the decrease in model inherent accuracy caused by the subset of target features compared to all sensor features, the preliminary fault probability is corrected to obtain a corrected warning probability.

[0124] This step aims to compensate for the systematic bias in the model caused by feature simplification. Because the lightweight early warning model is trained on a subset of target features, rather than all features, its inherent discriminative power differs from the original early warning model that uses all features. To obtain a probability estimate that more closely approximates the confidence level of the full-feature model, P needs to be adjusted. pre The calibration is performed. The specific implementation method for calibration is as follows: (1) Generation and distribution of correction factors: In the model deployment step (S200) at the remote control terminal, the correction factor g is calculated as a key parameter and encapsulated in the configuration information, and distributed to the local monitoring equipment along with the lightweight early warning model. This factor is directly related to the target feature subset F. target Performance.

[0125] (2) Basis for calculating the correction factor: The correction factor g is calculated by the remote control terminal based on the accuracy retention rate A(F) of the target feature subset. target A(F) is calculated to be... target This reflects the degree to which the model performance is preserved relative to the full-feature model when using a subset of the target features (its calculation method has been described in the joint optimization section of S103). A direct mapping relationship is: g = 1 / A(F target ).

[0126] When A(F) target The closer A(F) is to 1 (the smaller the performance loss), the closer g is to 1, and the smaller the correction magnitude; when A(F) is closer to 1, the correction magnitude is smaller. target The lower the value of g (the greater the performance loss), the greater the value of g, and the larger the correction magnitude.

[0127] (3) Correction calculation: The local monitoring equipment reads the pre-stored correction factor g and performs the following calculation on the preliminary fault probability to obtain the corrected early warning probability P. final :P final =min(1,P) pre ×g).

[0128] The min function is used to clamp the probability value within a reasonable range of [0, 1]. This calculation is equivalent to P final =P pre / A(F target Its physical meaning is to back-map the probability estimate obtained based on the subset model to the probability scale of the full feature model, thereby offsetting the model confidence decay caused by feature loss.

[0129] S303, generate the fault warning information based on the corrected warning probability.

[0130] The local monitoring equipment will correct the early warning probability P final It compares the results with multiple preset fault probability threshold ranges. For example, it can be set as follows: If P final <P1, judged as normal state; If P1≤P final <P2, indicating a warning status; If P final If the value is greater than or equal to P2, the system is considered to be in an alarm state. Here, P1 is the first fault probability threshold, which can be, for example, 0.3, and P2 is the second fault probability threshold, which can be, for example, 0.7.

[0131] Based on the comparison results, corresponding fault warning information is generated. This fault warning information can be directly displayed through indicator lights (such as green / yellow / red) on the local monitoring device, and can also be reported to the remote control terminal through the communication module.

[0132] Furthermore, the method also includes the following steps: S400, the remote control terminal optimizes and updates the target feature subset and the lightweight early warning model based on the continuously acquired operating data of the target device, and synchronizes the update results to the local monitoring device.

[0133] The remote control terminal continuously receives and accumulates operational data and associated working condition information from the target device. Based on the newly accumulated data, the currently used target feature subset and corresponding lightweight early warning model are iteratively optimized and updated, and the updated target feature subset and lightweight early warning model are sent to the local monitoring device to replace the original version.

[0134] The S400 is triggered by predefined conditions. The update process is automatically started when at least one of the following conditions is met: Data volume trigger: The newly accumulated operational data sample size of the target device reaches a preset update threshold. The preset threshold can be determined based on the model iteration update cycle requirements or the stability of the data distribution.

[0135] In one embodiment, the update threshold can be set as a proportion of the initial training sample size (e.g., 20%-50%) to ensure that the newly accumulated data is sufficient to reflect potential changes in equipment status or operating conditions, thereby supporting an effective model retraining. In another embodiment, the threshold can be determined based on statistical methods, for example, an update is triggered when the accumulation of new data causes a significant change in key statistical characteristics of the overall sample (such as mean, variance) (determined through hypothesis testing).

[0136] Performance feedback trigger: The remote control terminal receives confirmed false alarms or missed alarms in response to historical early warning information.

[0137] External knowledge trigger: The remote control terminal obtains new candidate features from the associated data platform or expert system, and confirms through statistical analysis or model verification that the candidate features have significant discriminative power for the fault mode of the current target device, that is, they contribute to improving the performance of the initial early warning model.

[0138] The statistical analysis method is as follows: calculate the data distribution of the new candidate feature under different states (such as normal, fault A, fault B). If the difference in its distribution under different states is statistically significant (for example, the p value obtained by ANOVA is less than 0.05), it can be preliminarily considered to have discriminative power.

[0139] The model is validated by adding the new candidate feature to the existing feature set and retraining a test model on a historical dataset or a specially constructed validation set. If the model shows a significant improvement in key performance metrics (such as accuracy, recall, and F1 score) exceeding a preset threshold (e.g., 1%) after adding the feature, then the feature is considered to have discriminative power.

[0140] In summary, the equipment fault early warning analysis method provided by the embodiments of the present invention has at least the following advantages: 1. Solved the fundamental contradiction between edge-side resources and performance: For local monitoring devices with strictly limited processing power, memory, and energy consumption, an innovative cloud-based collaborative feature screening and model lightweighting mechanism was used to adapt complex analysis models to run on the edge side while ensuring early warning accuracy, thus breaking through the resource bottleneck.

[0141] 2. Achieved dynamic adaptive optimal feature selection: By constructing a joint optimization objective function that integrates accuracy preservation rate and resource overhead, and introducing a dynamic trade-off coefficient, the system can intelligently filter out the optimal feature subset of resource constraints at the edge based on the real-time resource status of the device, achieving the best balance between accuracy and efficiency.

[0142] 3. Deeply integrate domain knowledge into automated decision-making: By setting seed feature groups and differentiated judgment thresholds (or embedding priority penalty terms in the optimization function), the prior knowledge of domain experts is solidified into computable rules, ensuring a high retention priority for key fault features and improving the interpretability and reliability of the screening results.

[0143] 4. A complete closed loop of training-deployment-evolution has been constructed: Through continuous data accumulation and model update mechanism in the cloud, the system can automatically optimize the feature set and early warning model by using new data and feedback, so that the early warning capability can continuously evolve with the operation of the equipment, effectively cope with the degradation of equipment performance and the emergence of new failure modes, and realize the long-term effectiveness and intelligence of the predictive maintenance system.

[0144] 5. Improved engineering usability and accuracy of early warning results: By introducing a probability correction step based on feature selection performance into the local early warning process, the systematic bias of the model caused by feature simplification is compensated, making the final output early warning probability closer to the confidence level of the full feature model, thus improving the accuracy of decision-making.

[0145] This invention also provides an electronic device, including: 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 configured to perform the method described in this invention.

[0146] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0147] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0148] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for early warning analysis of equipment faults, characterized in that, A method applicable to a system comprising a remote control terminal and at least one local monitoring device, wherein each local monitoring device is connected to at least one target device, the method comprising the following steps: S100: The remote control terminal performs feature filtering based on the initial early warning model trained for the target device to obtain a subset of target features; S200, the remote control terminal generates a corresponding lightweight early warning model based on the target feature subset and sends it to the corresponding local monitoring device; S300, the local monitoring device acquires the monitoring data of the target device corresponding to the target feature subset, processes it using the lightweight early warning model, and generates and outputs fault early warning information for the target device.

2. The method according to claim 1, characterized in that, In S100, the feature selection is achieved through the following steps: S101, divide all sensor features of the initial early warning model into multiple feature groups; wherein, the feature groups include at least a seed feature group defined based on prior knowledge; S102, for the multiple feature groups, the importance measure of each feature group to the performance of the initial warning model is determined by evaluating the decrease in accuracy of the output of the initial warning model when different feature groups are masked in the input of the initial warning model. S103, based on the importance metric, and on the premise of meeting the resource constraints of the local monitoring equipment, select one or more feature groups that make the accuracy decrease of the initial early warning model within a preset cumulative accuracy decrease threshold to form the target feature subset.

3. The method according to claim 2, characterized in that, The precision reduction value for each feature group is determined through the following steps: In the input data of the initial early warning model, all feature data belonging to the feature group are masked, and the masked data is input into the initial early warning model to obtain the corresponding evaluation accuracy; The difference between the evaluation accuracy and the baseline accuracy of the initial early warning model when all feature data are input is calculated. This difference is the accuracy reduction value corresponding to the feature group.

4. The method according to claim 2, characterized in that, In S103, the selection of the target feature subset F is achieved by constructing and maximizing a joint optimization objective function Q, where, Q(F)=α·A(F)-(1-α)·C(F); Where A(F) represents the accuracy retention rate of the initial early warning model when using a subset of candidate features F; C(F) represents the normalized comprehensive overhead of collecting and processing F on the local monitoring device; and α is a trade-off coefficient dynamically adjusted according to the real-time resource status of the local monitoring device.

5. The method according to claim 2, characterized in that, For feature group G, the importance metric I(G) of feature group G is β·ΔAccuracy(G) + (1-β)·ΔSeparability(G); where ΔAccuracy(G) represents the decrease in accuracy of the initial warning model caused by masking feature group G, ΔSeparability(G) represents the degree of decrease in the distinguishability of the feature representation of different fault modes of the initial warning model caused by masking feature group G, and β is a preset fusion weight parameter.

6. The method according to claim 1, characterized in that, The process of using the lightweight early warning model to generate and output fault early warning information for the target device specifically includes: S301, the lightweight early warning model calculates a preliminary failure probability based on the input monitoring data; S302, Based on the decrease in model inherent accuracy caused by the target feature subset compared to all sensor features, the preliminary fault probability is corrected to obtain a corrected warning probability; S303, generate the fault warning information based on the corrected warning probability.

7. The method according to claim 1, characterized in that, The method further includes: S400, the remote control terminal optimizes and updates the target feature subset and the lightweight early warning model based on the continuously acquired operating data of the target device, and synchronizes the update results to the local monitoring device.

8. The method according to claim 7, characterized in that, S400 is triggered when at least one of the following conditions is met: The amount of accumulated operating data of the target device has reached a preset update threshold; The remote control terminal receives confirmed feedback regarding false alarms or missed alarms related to historical early warning information; The remote control terminal acquires new features that, after evaluation, contribute to improving the performance of the initial early warning model.

9. An electronic device, characterized in that, Including processor and memory; The processor executes the steps of the method as described in any one of claims 1 to 8 by invoking programs or instructions stored in the memory.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a program or instructions that cause a computer to perform the steps of the method as described in any one of claims 1 to 8.

Images