Artificial intelligence industrial equipment state detection method

Abstract

The invention discloses an artificial intelligence industrial equipment state detection method, and relates to the technical field of industrial equipment detection, and the method comprises the following steps: obtaining time sequence operation data of each operation key part of industrial equipment, carrying out the framing processing of the time sequence operation data, and obtaining a plurality of groups of frame data, extracting feature vectors of each group of frame data to construct an equipment operation feature matrix; performing hierarchical feature extraction on the equipment operation feature matrix based on the lightweight convolution kernel to obtain a shallow time domain feature and a deep frequency domain feature, and fusing the shallow time domain feature and the deep frequency domain feature to obtain a fused feature vector; the method has the advantages that the limitation of a single scene scheme is solved, the requirement for full-life-cycle management of industrial equipment can be met, the equipment failure outage rate is effectively reduced, and the continuity and safety of industrial production are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment inspection technology, and more specifically, to an artificial intelligence-based method for industrial equipment condition detection. Background Technology

[0002] As the core carrier of modern industrial production, the stability of industrial equipment's operating status directly determines production efficiency, product quality, and operational safety. With the advancement of Industry 4.0 and intelligent manufacturing, industrial equipment is developing towards larger scale, continuous operation, and intelligence. The number of operating components in the equipment is increasing, and the correlation between parameters is significantly improving. Traditional status detection methods such as manual inspection and fixed-point monitoring can no longer meet the requirements for real-time, accurate, and comprehensive detection.

[0003] Convolution, as a core tool in the field of artificial intelligence, has demonstrated powerful feature extraction capabilities in scenarios such as image recognition and natural language processing, and is gradually extending to the field of industrial production. However, existing industrial equipment condition detection technologies mostly use a holistic feature extraction approach to process equipment time-series operation data, without performing fine-grained frame-by-frame processing on the data. This easily leads to the loss of local operational features, resulting in insufficient accuracy in feature representation. At the same time, feature extraction often involves extracting only time-domain or frequency-domain features, without achieving deep fusion of time-domain and frequency-domain features, thus failing to comprehensively represent the operating status of the equipment. Furthermore, fault matching simply calculates feature similarity without combining fault levels to quantify the probability of anomalies, resulting in low accuracy in anomaly detection and an inability to achieve precise reverse localization of abnormal parameters.

[0004] In specific scenarios such as parts quality inspection, existing technologies also face challenges in implementation. Complex model optimization schemes based on image features, or approaches that improve performance through network structure, are not only costly to train but also demanding in terms of hardware resources, making them unaffordable for small factories. To address this pain point, some lightweight convolutional solutions have been proposed; however, these solutions mostly focus on single quality inspection scenarios and have not formed a complete technology system covering equipment status detection, fault location, and maintenance optimization, thus failing to meet the needs of full lifecycle management of industrial equipment. Summary of the Invention

[0005] In view of the shortcomings of existing technologies, the purpose of this invention is to provide an artificial intelligence-based method for industrial equipment condition detection.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An artificial intelligence-based method for monitoring the condition of industrial equipment, comprising the following steps: Acquire the time-series operation data of each key component of the industrial equipment, perform frame segmentation on the time-series operation data to obtain multiple sets of frame data, and extract the feature vector of each set of frame data to construct the equipment operation feature matrix; Based on lightweight convolutional kernels, the device operation feature matrix is ​​subjected to hierarchical feature extraction to obtain shallow time-domain features and deep frequency-domain features. The shallow time-domain features and deep frequency-domain features are then fused to obtain a fused feature vector. Extract a fault feature template library of similar equipment from the historical equipment fault database, and calculate the similarity matching value between the fused feature vector and each template in the fault feature template library; The probability of an anomaly in the device's operating status is determined based on the similarity matching value. If the probability of an anomaly is greater than a preset probability threshold, the key operating parameters associated with the device anomaly are located in reverse based on the fused feature vector. Based on key operating parameters, a preliminary adjustment plan for the equipment status is generated. The execution priority of the preliminary adjustment plan is dynamically optimized in combination with the equipment operating load, and the target maintenance plan is output.

[0007] Preferably, the process involves acquiring time-series operational data of key components of the industrial equipment, performing frame-segmentation on the time-series operational data to obtain multiple sets of frame data, and extracting the feature vectors of each set of frame data to construct an equipment operational feature matrix, including: The system collects operating data of key components of the equipment, such as motors, bearings, and hydraulic systems, using sensors. The operating data includes vibration amplitude, temperature changes, current fluctuations, and time-series change curves. The time-series running data is processed into frames according to a preset time window, and multiple sets of frame data are obtained by setting a preset frame overlap rate threshold. Extract the time-domain statistical features and frequency-domain features of each set of frame data. The time-domain statistical features include mean, variance, and peak factor. The frequency-domain features include the main frequency amplitude and harmonic components. The time-domain statistical features and frequency-domain features of each set of frame data are combined into a feature vector, and then arranged in frame order to construct a device operation feature matrix.

[0008] Preferably, a layered feature extraction method is used to extract shallow time-domain features and deep frequency-domain features from the device operation feature matrix based on lightweight convolutional kernels. The shallow time-domain features and deep frequency-domain features are then fused to obtain a fused feature vector, including: Construct a first lightweight convolutional kernel and a second lightweight convolutional kernel. The first lightweight convolutional kernel is used to extract local features in the temporal domain, and the second lightweight convolutional kernel is used to extract correlation features in the frequency domain. Based on the first lightweight convolutional kernel, shallow temporal features are obtained by calculating the device operation feature matrix. Based on the second lightweight convolution kernel, the shallow time-domain features are calculated to obtain the deep frequency-domain features; Dynamic weights are assigned to shallow time-domain features and deep frequency-domain features to obtain a fused feature vector.

[0009] Preferably, a fault feature template library for similar equipment is extracted from a historical equipment fault database, and the similarity matching value between the fused feature vector and each template in the fault feature template library is calculated, including: Filter fault records that match the target equipment model and operating conditions from the historical equipment fault database, and extract the fault feature vectors corresponding to each fault record as fault feature templates; All fault feature templates are categorized and organized according to fault type to build a fault feature template library. Fault types include mechanical wear, electrical faults, and hydraulic leakage. Calculate the similarity between the fused feature vector and each template in the fault feature template library to obtain multiple similarity matching values; Multiple similarity matching values ​​are filtered, and those with similarity matching values ​​greater than a preset threshold are marked as valid matching results.

[0010] Preferably, the probability of an anomaly in the device's operating status is determined based on the similarity matching value. If the anomaly probability is greater than a preset probability threshold, key operating parameters associated with the device anomaly are located based on the fused feature vector, including: The anomaly probability is calculated from the similarity matching values ​​in the valid matching results; If the probability of an anomaly is less than the preset probability threshold, the device is determined not to have any operational anomalies. If the probability of abnormality is greater than or equal to the preset probability threshold, the device is determined to have an operational abnormality. The contributing features are obtained by selecting the feature dimension that contributes the most to the anomaly probability from the fused feature vector; The contribution features are mapped to the corresponding device operating parameters, which are then identified as key operating parameters associated with anomalies.

[0011] Preferably, a pre-selected adjustment plan for equipment status is generated based on key operating parameters, and the execution priority of the pre-selected adjustment plan is dynamically optimized in conjunction with the equipment operating load to output a target maintenance plan, including: For each key operating parameter, a corresponding pre-selected adjustment plan is generated based on historical optimization data. The pre-selected adjustment plan includes the parameter adjustment range and adjustment steps. Collect current operating load data of the equipment, including load rate, continuous running time, and urgency of production tasks; Based on the operational load data and the impact range of the pre-selected adjustment schemes, the priority score of each pre-selected adjustment scheme is calculated; the pre-selected adjustment schemes are sorted from high to low according to their priority scores, integrated to form the target maintenance scheme, and the target maintenance scheme is output.

[0012] Preferably, it further includes: The effectiveness of the target maintenance plan is monitored in real time, equipment operation data after the plan is implemented is collected, and new fusion feature vectors are extracted. Calculate the similarity matching value between the new fused feature vector and the fault feature template library to determine the probability of anomalies after the solution is implemented; If the probability of anomalies after the implementation of the solution is less than the preset probability threshold, the maintenance is deemed effective; if the probability of anomalies is greater than or equal to the preset probability threshold, the fault feature template library is updated and a pre-selected adjustment solution for the equipment status is regenerated.

[0013] Compared with existing technologies, this invention has the following advantages: It segments time-series operational data using an overlapping framing method, avoiding the loss of local features; simultaneously, it extracts shallow time-domain features and deep frequency-domain features layer by layer, and then uses a weighted fusion algorithm to organically combine multi-dimensional features, preserving both the dynamic time-domain information of equipment operation and capturing fault-related features in the frequency domain, significantly improving the accuracy and comprehensiveness of feature representation; by combining the real-time operating load of the equipment, it constructs correlation parameters between load and execution cost, and operation time, dynamically optimizing the execution priority of maintenance plans. During high-load operation, it prioritizes low-cost, short-time adjustment plans, and during low-load operation, it prioritizes root-cause solutions, ensuring that maintenance plans fully adapt to production conditions and avoiding disruption to production continuity due to improper maintenance operations, thus significantly improving the practicality and operability of maintenance plans; through the entire process of data acquisition, feature extraction, anomaly detection, fault location, maintenance plan generation, and effect verification, it constructs a closed-loop verification mechanism for maintenance effects, overcoming the limitations of single-scenario solutions, meeting the needs of full lifecycle management of industrial equipment, effectively reducing equipment downtime rates, and ensuring the continuity and safety of industrial production. Attached Figure Description

[0014] Figure 1 This invention provides a schematic diagram of the steps of an artificial intelligence-based industrial equipment condition detection method according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the steps of constructing a device operation feature matrix in an artificial intelligence-based industrial equipment condition detection method according to an embodiment of the present invention. Detailed Implementation

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0016] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0017] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0018] Reference Figures 1-2 As shown.

[0019] This embodiment further illustrates the artificial intelligence-based industrial equipment condition detection method proposed in this invention.

[0020] An artificial intelligence-based method for monitoring the condition of industrial equipment, comprising the following steps: Acquire the time-series operation data of each key component of the industrial equipment, perform frame segmentation on the time-series operation data to obtain multiple sets of frame data, and extract the feature vector of each set of frame data to construct the equipment operation feature matrix; Based on lightweight convolutional kernels, the device operation feature matrix is ​​subjected to hierarchical feature extraction to obtain shallow time-domain features and deep frequency-domain features. The shallow time-domain features and deep frequency-domain features are then fused to obtain a fused feature vector. Extract a fault feature template library of similar equipment from the historical equipment fault database, and calculate the similarity matching value between the fused feature vector and each template in the fault feature template library; The probability of an anomaly in the device's operating status is determined based on the similarity matching value. If the probability of an anomaly is greater than a preset probability threshold, the key operating parameters associated with the device anomaly are located in reverse based on the fused feature vector. Based on key operating parameters, a preliminary adjustment plan for the equipment status is generated. The execution priority of the preliminary adjustment plan is dynamically optimized in combination with the equipment operating load, and the target maintenance plan is output.

[0021] Acquire time-series operational data of key components of industrial equipment, perform frame-segmentation processing on the time-series operational data to obtain multiple sets of frame data, extract feature vectors from each set of frame data to construct an equipment operation feature matrix, including: The system collects operating data of key components of the equipment, such as motors, bearings, and hydraulic systems, using sensors. The operating data includes vibration amplitude, temperature changes, current fluctuations, and time-series change curves. The time-series running data is processed into frames according to a preset time window, and multiple sets of frame data are obtained by setting a preset frame overlap rate threshold. Extract the time-domain statistical features and frequency-domain features of each set of frame data. The time-domain statistical features include mean, variance, and peak factor. The frequency-domain features include the main frequency amplitude and harmonic components. The time-domain statistical features and frequency-domain features of each set of frame data are combined into a feature vector, and then arranged in frame order to construct a device operation feature matrix.

[0022] First, the system continuously collects operational data using various sensors deployed at key parts of the equipment. Taking an industrial pump as an example, the sensors collect data on motor current fluctuations, bearing vibration amplitudes, and hydraulic system temperature changes, forming continuous time-series curves. This real-time data directly reflects the dynamic operating status of each component of the equipment and serves as the primary basis for subsequent analysis.

[0023] Next, the system will perform frame segmentation on the continuous time-series data. This breaks down the long time series into multiple locally representative data segments, facilitating the capture of transient anomalies during equipment operation. Taking bearing vibration data as an example, if the time window length is set to 1 second and the frame overlap rate is 40%, a new data frame will be generated every 600 milliseconds, ensuring that adjacent frames have a 40% overlap. This processing method not only breaks down long time-series data into local segments but also avoids the loss of transient anomalies due to data truncation, preserving complete local information for subsequent feature extraction.

[0024] After frame segmentation, the system performs deep feature extraction on each set of frame data, uncovering the inherent patterns in the data from both the time and frequency domains. In the time-domain feature extraction stage, the system calculates the mean, variance, and peak value. The mean is the arithmetic average of the data in a given frame, calculated using the following formula: , where x i Let be the value of the i-th sampling point within the frame, and n be the number of sampling points within the frame. It reflects the average level of the parameter within that time period; for example, the average value of the motor current in a certain frame can reflect the average load during that period. The formula for calculating variance is: It reflects the degree of data fluctuation; for example, a large current variance indicates unstable fluctuations in the motor load. The formula for calculating the peak factor is: It is the ratio of peak value to root mean square value within a frame, and it can capture instantaneous impact signals. For example, when pitting occurs in a bearing, the peak factor of the vibration data will increase significantly.

[0025] In the frequency domain feature extraction stage, the system first performs a Fourier transform on the intra-frame time-domain data to convert it into a frequency-domain signal. The dominant frequency amplitude is the amplitude corresponding to the highest energy frequency in the frequency-domain signal; it reflects the dominant vibration frequency of the equipment. For example, the dominant frequency amplitude of a normal bearing is stable, but when the bearing wears, the amplitude of the dominant frequency at a specific frequency will increase significantly. Harmonic components are the harmonics of the dominant frequency signal, and their calculation formula is: Where X(f0) is the amplitude of the frequency domain signal at frequency f, and f0 is the main frequency. Abnormal harmonic components often point to problems such as poor gear meshing or motor winding faults.

[0026] Finally, the system integrates the temporal and frequency domain statistical features of each frame into a single feature vector. Then, it arranges the feature vectors corresponding to all frames in chronological order, ultimately constructing a device operation feature matrix. Each row of the device operation feature matrix represents the multi-dimensional features of a frame of data, and each column corresponds to a specific feature type, forming a dynamic map of the device's operating status. This comprehensively depicts the device's operating characteristics at different time segments, providing accurate and systematic data support for subsequent AI-based fault diagnosis and status assessment.

[0027] Based on lightweight convolutional kernels, hierarchical feature extraction is performed on the device operation feature matrix to obtain shallow time-domain features and deep frequency-domain features. The shallow time-domain features and deep frequency-domain features are then fused to obtain a fused feature vector, including: Construct a first lightweight convolutional kernel and a second lightweight convolutional kernel. The first lightweight convolutional kernel is used to extract local features in the temporal domain, and the second lightweight convolutional kernel is used to extract correlation features in the frequency domain. Based on the first lightweight convolutional kernel, shallow temporal features are obtained by calculating the device operation feature matrix. Based on the second lightweight convolution kernel, the shallow time-domain features are calculated to obtain the deep frequency-domain features; Dynamic weights are assigned to shallow time-domain features and deep frequency-domain features to obtain a fused feature vector.

[0028] First, the system constructs two types of lightweight convolutional kernels with clearly defined functions. The first type employs a depthwise separable convolution structure to capture local, continuous variations in the time domain. This type of kernel decomposes standard convolution into depthwise convolution and pointwise convolution. Taking the operating feature matrix of an industrial pump as an example, depthwise convolution convolves each input channel of the operating feature matrix individually, capturing the changing trend of bearing vibration amplitude within adjacent time frames. Pointwise convolution then fuses information between channels in the output of the depthwise convolution, ultimately obtaining an intermediate result containing local time-domain features. The second type of lightweight convolutional kernel combines pointwise convolution with Fourier transform to further extract frequency-domain correlation features from the time-domain features. This second type of kernel transforms the channel dimensions of shallow time-domain features through pointwise convolution, and then converts the time-domain signal into a frequency-domain signal through Fourier transform, thereby capturing the main frequency changes and harmonic correlations during equipment operation.

[0029] Next, the system calculates the device operation feature matrix based on the first lightweight convolutional kernel to obtain shallow temporal features. Taking the 1439×12 operation feature matrix of an industrial pump as an example, the kernel size of the first lightweight convolutional kernel is set to 3×1 with a stride of 1. It slides frame by frame, performing local convolution calculations on the features of every 3 consecutive time frames, outputting a shallow temporal feature matrix with a dimension of 1437×8. Each element in this matrix represents the temporal variation law of the device operation parameters within the corresponding time segment, such as the current fluctuation trend and continuous change of vibration amplitude in a certain frame. These features directly reflect the dynamic operating state of the device in the time dimension and are the basis for subsequent frequency domain analysis.

[0030] Subsequently, the shallow time-domain features are processed using a second lightweight convolutional kernel to obtain deep frequency-domain features. The second lightweight convolutional kernel first transforms the channel dimension of the shallow time-domain features from 8 to 16 through a 1×8 point convolution, and then performs a Fast Fourier Transform on the transformed features to convert the time-domain signal into a frequency-domain signal. Taking the bearing vibration data of an industrial pump as an example, the frequency domain transformation yields the characteristics of the dominant frequency amplitude and harmonic components. The dominant frequency amplitude reflects the dominant vibration frequency of the bearing operation; when the bearing wears, the dominant frequency amplitude at a specific frequency will increase significantly. The harmonic components are the harmonics of the dominant frequency; abnormal harmonic components often indicate poor gear meshing or motor winding faults. Finally, this step outputs a deep frequency-domain feature matrix with a dimension of 1437×16, comprehensively characterizing the frequency domain features of the equipment operation.

[0031] Finally, the system dynamically assigns weights to the shallow time-domain features and the deep frequency-domain features and then fuses them. The dynamic weight allocation is based on historical data of similar faults. For example, in historical cases of industrial pump bearing wear, frequency-domain features contribute more to fault diagnosis; therefore, a weight of 0.55 is assigned to the frequency-domain features, and a weight of 0.45 is assigned to the time-domain features. The fusion calculation formula is as follows: In this fusion, F represents the fused feature vector, T represents the shallow time-domain feature, Ff represents the deep frequency-domain feature, and α and β are the weighting coefficients of the time-domain and frequency-domain features, respectively, satisfying α+β=1. Through this weighted fusion, the system can organically combine the dynamic change information in the time domain with the frequency correlation information in the frequency domain, forming a fused feature vector that can more comprehensively reflect the operating status of the equipment. Taking an industrial pump as an example, the fused feature vector includes both the time-domain trend of current fluctuations and the frequency-domain anomalies of vibration signals, enabling more accurate identification of early bearing wear faults and providing a solid feature foundation for subsequent anomaly judgment and maintenance scheme generation.

[0032] Extract a fault feature template library of similar equipment from a historical equipment fault database, and calculate the similarity matching value between the fused feature vector and each template in the fault feature template library, including: Filter fault records that match the target equipment model and operating conditions from the historical equipment fault database, and extract the fault feature vectors corresponding to each fault record as fault feature templates; All fault feature templates are categorized and organized according to fault type to build a fault feature template library. Fault types include mechanical wear, electrical faults, and hydraulic leakage. Calculate the similarity between the fused feature vector and each template in the fault feature template library to obtain multiple similarity matching values; Multiple similarity matching values ​​are filtered, and those with similarity matching values ​​greater than a preset threshold are marked as valid matching results.

[0033] First, the system filters out fault records from the historical equipment fault database that are completely consistent with the target equipment model and operating conditions. Taking an industrial pump as an example, if the target equipment is a centrifugal industrial pump of a certain model, operating under the condition of conveying room temperature clean water with a rated pressure of 0.8MPa, the system will search the database for all records of faults that have occurred under the same model and operating conditions, and extract the fused feature vectors corresponding to these faults as fault feature templates. These templates contain typical features of different fault types such as mechanical wear, electrical faults, and hydraulic leaks under corresponding operating conditions. For example, the mechanical wear template will include the time domain features of abnormal bearing vibration frequency and temperature rise; the electrical fault template will include the frequency domain features of motor current fluctuations and harmonic distortion; and the hydraulic leak template will include the time domain features of hydraulic system pressure drop and flow fluctuations.

[0034] Next, the system will categorize and organize all fault feature templates according to fault type, constructing a structured fault feature template library. For example, the mechanical wear template will include sub-templates for bearing wear and impeller wear; the electrical fault template will include sub-templates for motor overload and winding insulation degradation; and the hydraulic leakage template will include sub-templates for seal damage and pipeline cracking. Each sub-template corresponds to a typical fusion feature vector of a specific fault. This categorization allows subsequent similarity matching to be more targeted, improving the efficiency and accuracy of fault identification.

[0035] Subsequently, the system calculates the similarity between the fused feature vector and each template in the fault feature template library, obtaining multiple similarity matching values. The most commonly used algorithm here is cosine similarity, and its calculation formula is as follows: In this model, F is the fused feature vector calculated in real time, M is a fault feature template from the template library, F·M is the dot product of the two, and ||F|| and ||M|| are the magnitudes of the two vectors, respectively. Taking the fused feature vector of an industrial pump as an example, if the cosine similarity of this vector with the bearing wear template of the mechanical wear category is 0.89, the similarity with the motor overload template of the electrical fault category is 0.32, and the similarity with the seal damage template of the hydraulic leakage category is 0.21, this indicates that the current operating state of the equipment is most similar to the typical characteristics of bearing wear.

[0036] Finally, the system filters multiple similarity matches, marking those with similarity values ​​greater than a preset threshold as valid matches. For example, if the preset threshold is set to 0.7, then the similarity match value of 0.89 in the example above will be marked as a valid match, while 0.32 and 0.21 will be judged as invalid. Valid matches directly point to the corresponding fault type, providing a clear direction for subsequent anomaly probability calculation and fault location. For instance, a valid match might prompt the system to focus on bearing-related operating parameters to further verify whether the equipment has a bearing wear fault.

[0037] The probability of an anomaly in the device's operating status is determined based on the similarity matching value. If the anomaly probability is greater than a preset probability threshold, key operating parameters associated with the device anomaly are located based on the fused feature vector, including: The anomaly probability is calculated from the similarity matching values ​​in the valid matching results; If the probability of an anomaly is less than the preset probability threshold, the device is determined not to have any operational anomalies. If the probability of abnormality is greater than or equal to the preset probability threshold, the device is determined to have an operational abnormality. The contributing features are obtained by selecting the feature dimension that contributes the most to the anomaly probability from the fused feature vector; The contribution features are mapped to the corresponding device operating parameters, which are then identified as key operating parameters associated with anomalies.

[0038] After the system completes the similarity matching between the fused feature vector and the fault template library, it first performs a weighted calculation on the similarity matching values ​​among all valid matching results to obtain the probability of an anomaly in the device's operating status. The formula for calculating the anomaly probability is: , where P raw For the anomaly probability, cosθ i w is the cosine similarity value of the i-th valid match. iLet k be the fault level coefficient of the fault template corresponding to the matching result, and k be the number of valid matching results. For example, in the detection scenario of industrial pumps, if the only valid matching result is the bearing wear template, its similarity is 0.89, and the corresponding fault level coefficient is 2. Substituting this into the formula, the anomaly probability can be calculated as 0.89 × 2 = 1.78. After normalization, the anomaly probability value is obtained. The core of normalization is to map the initial anomaly probability to the range of 0% to 100%. The specific formula is as follows: , where P norm P is the normalized outlier probability. max This represents the maximum initial anomaly probability for similar equipment in historical data. For example, in the case of this industrial pump, according to the historical fault database, the maximum initial anomaly probability for similar equipment is 2.04. Therefore, substituting this into the calculation yields... The calculated probability of an anomaly is 87.2%.

[0039] After obtaining the anomaly probability, the system compares it with a preset probability threshold. If the preset probability threshold is 80%, and the calculated 87.2% is greater than the threshold, the system determines that the device is malfunctioning; if the anomaly probability is lower than the threshold, the system determines that the device is operating normally.

[0040] After determining that the equipment is abnormal, the system will further locate the key operating parameters associated with the anomaly. First, the system will perform a contribution analysis on each feature dimension in the fused feature vector, and select the feature dimensions that contribute the most to the anomaly probability, i.e., contributing features. For example, in the case of industrial pumps, the system will calculate the contribution of time-domain features and frequency-domain features in the fused feature vector, and find that the frequency-domain feature of bearing vibration main frequency amplitude contributes 42% to the anomaly probability, the time-domain feature of impeller speed fluctuation contributes 28%, and the motor current fluctuation contributes 18%. These feature dimensions with the highest contribution will be marked as contributing features.

[0041] Finally, the system maps these contributing features to corresponding equipment operating parameters. For example, the main frequency amplitude of bearing vibration corresponds to the bearing temperature and vibration amplitude parameters, impeller speed fluctuation corresponds to the impeller speed parameters, and motor current fluctuation corresponds to the motor current parameters. Through this mapping relationship, the system can accurately determine that bearing temperature, impeller speed, and motor current are key operating parameters associated with anomalies. These parameters directly point to the root cause of equipment anomalies, providing a clear basis for generating subsequent maintenance plans. For example, regarding the problem of increased bearing temperature, the system can prioritize proposing adjustment plans for bearing repair or lubrication optimization.

[0042] Based on key operating parameters, a preliminary adjustment plan for equipment status is generated. The execution priority of this plan is dynamically optimized in conjunction with the equipment operating load, and a target maintenance plan is output, including: For each key operating parameter, a corresponding pre-selected adjustment plan is generated based on historical optimization data. The pre-selected adjustment plan includes the parameter adjustment range and adjustment steps. Collect current operating load data of the equipment, including load rate, continuous running time, and urgency of production tasks; Based on the operational load data and the impact range of the pre-selected adjustment schemes, the priority score of each pre-selected adjustment scheme is calculated; the pre-selected adjustment schemes are sorted from high to low according to their priority scores, integrated to form the target maintenance scheme, and the target maintenance scheme is output.

[0043] Taking industrial pumps as an example, once the system identifies bearing temperature, impeller speed, and motor current as critical operating parameters with abnormal correlations, it will first generate corresponding pre-selected adjustment plans for each parameter. For instance, regarding rising bearing temperature, the system will generate a plan based on historical optimization data, including parameter adjustment ranges and specific steps. This might involve adjusting the bearing temperature warning threshold from 85℃ to 80℃, increasing the temperature inspection frequency to once every two hours, and outlining maintenance procedures for disassembling, cleaning, and replenishing lubricant on the bearing after shutdown. For impeller speed fluctuations, the plan will propose stabilizing the speed within ±2% of the rated value, achieved by calibrating the inverter output parameters. For motor current fluctuations, the plan will suggest adjusting the pump's outlet valve opening to control the current within 90% of the rated current, avoiding overload.

[0044] Next, the system will collect the current operating load data of the equipment, including load rate, continuous running time, and the urgency of the production task. In the scenario of industrial pumps, if the current equipment load rate is 95%, the continuous running time has reached 12 hours, and the production task is in a critical stage of conveying raw materials, with a high urgency level, this means that the equipment cannot be shut down for maintenance temporarily, otherwise it will directly affect the subsequent production process.

[0045] Then, based on the operating load data and the impact range of the pre-selected adjustment schemes, the system calculates the priority score for each pre-selected adjustment scheme. The formula for calculating the priority score is: Where S is the priority score, L is the load rate, T is the continuous running time, E is the urgency of the production task, C is the execution cost and time coefficient of the solution, and a, b, c, and d are weighting coefficients. In this industrial pump case, the weighting coefficients can be set as a=0.3, b=0.2, c=0.3, and d=0.2. If the execution cost and time coefficient of the valve opening adjustment solution is 0.1, the coefficient for calibrating the frequency converter is 0.3, and the coefficient for bearing disassembly and inspection is 0.8, substituting the data, the priority score for adjusting the valve opening can be calculated as: S1=0.3×95+0.2×12+0.3×10 0.2 × 0.1 ≈ 32.58; Priority score for calibrating the frequency converter: S2 = 0.3 × 95 + 0.2 × 12 + 0.3 × 10 0.2 × 0.3 ≈ 32.54; Priority score for bearing disassembly and inspection: S3 = 0.3 × 95 + 0.2 × 12 + 0.3 × 10 0.2 × 0.8 ≈ 32.44; The system will prioritize and pre-select adjustment schemes according to their scores, from highest to lowest, and integrate them to form a target maintenance plan. In this case, the highest-scoring scheme, adjusting the valve opening, will be set as the first-level execution, calibrating the frequency converter as the second-level execution, and disassembling and inspecting the bearing as the third-level execution. The final output target maintenance plan will clearly state that adjusting the valve opening will be executed first to quickly reduce the motor current. After the production task enters a non-critical stage and the equipment load rate decreases, the frequency converter calibration and bearing disassembly and inspection will be executed sequentially. The operational requirements and monitoring indicators for each step will be clearly marked to ensure that the maintenance operation resolves potential faults without affecting production continuity.

[0046] Also includes: The effectiveness of the target maintenance plan is monitored in real time, equipment operation data after the plan is implemented is collected, and new fusion feature vectors are extracted. Calculate the similarity matching value between the new fused feature vector and the fault feature template library to determine the probability of anomalies after the solution is implemented; If the probability of anomalies after the implementation of the solution is less than the preset probability threshold, the maintenance is deemed effective; if the probability of anomalies is greater than or equal to the preset probability threshold, the fault feature template library is updated and a pre-selected adjustment solution for the equipment status is regenerated.

[0047] Taking the maintenance of industrial pumps as an example, once the system outputs and executes the target maintenance plan, it immediately initiates real-time monitoring of the maintenance effect. The system continuously collects post-maintenance equipment operation data through sensors, including time-series variation curves of key parameters such as bearing temperature, impeller speed, and motor current. Following the previous frame-by-frame processing, feature extraction, and fusion workflow, a new fused feature vector is generated. This new fused feature vector contains the time-domain and frequency-domain features of the post-maintenance equipment operating status and serves as the core data basis for verifying the maintenance effect.

[0048] Next, the system will perform similarity matching between the newly fused feature vector and all templates in the fault feature template library to calculate a new similarity matching value. Then, through weighted calculation and normalization, the anomaly probability after maintenance will be obtained. Using the previous calculation formula, if the similarity between the newly fused feature vector after maintenance and the bearing wear template is 0.21, and the corresponding fault level coefficient is 2, then the initial anomaly probability is P. raw =0.21×2=0.42, and considering the historical maximum initial anomaly probability of 2.04 for similar equipment, the normalized anomaly probability is P. norm=0.42 / 2.04×100%≈20.6%. Comparing this result with the preset 80% probability threshold, 20.6% is far below the threshold, so the system will determine that the maintenance is effective and the equipment's operating status has been restored to the normal range.

[0049] If the anomaly probability after maintenance is still greater than or equal to a preset threshold—for example, if the similarity between the newly fused feature vector and the bearing wear template after maintenance is 0.78, and the calculated normalized anomaly probability is 76.5%, close to the 80% threshold—the system determines that the maintenance has not completely resolved the fault. In this case, the system will supplement the fault feature template library with the newly fused feature vector and fault data collected during the maintenance process, update the feature dimensions of the original templates, or add fault templates adapted to the current operating conditions. Based on the updated template library, it will then regenerate pre-selected adjustment schemes, such as adding bearing clearance detection or changing the type of lubricating grease, forming a new maintenance loop. This ensures that the system can continuously learn the fault evolution patterns of the equipment, continuously optimize the accuracy of the fault feature template library and maintenance schemes, and enable the entire equipment condition monitoring and maintenance system to have dynamic iteration capabilities, adapting to long-term changes in equipment operating conditions.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-based method for industrial equipment condition detection, characterized in that, The method includes the following steps: Acquire the time-series operation data of each key component of the industrial equipment, perform frame segmentation on the time-series operation data to obtain multiple sets of frame data, and extract the feature vector of each set of frame data to construct the equipment operation feature matrix; Based on lightweight convolutional kernels, the device operation feature matrix is ​​subjected to hierarchical feature extraction to obtain shallow time-domain features and deep frequency-domain features. The shallow time-domain features and deep frequency-domain features are then fused to obtain a fused feature vector. Extract a fault feature template library of similar equipment from the historical equipment fault database, and calculate the similarity matching value between the fused feature vector and each template in the fault feature template library; The probability of an anomaly in the device's operating status is determined based on the similarity matching value. If the probability of an anomaly is greater than a preset probability threshold, the key operating parameters associated with the device anomaly are located in reverse based on the fused feature vector. Based on key operating parameters, a preliminary adjustment plan for the equipment status is generated. The execution priority of the preliminary adjustment plan is dynamically optimized in combination with the equipment operating load, and the target maintenance plan is output.

2. The artificial intelligence-based industrial equipment condition detection method according to claim 1, characterized in that, Acquire time-series operational data of key components of industrial equipment, perform frame-segmentation processing on the time-series operational data to obtain multiple sets of frame data, extract feature vectors from each set of frame data to construct an equipment operation feature matrix, including: The system collects operating data of key components of the equipment, such as motors, bearings, and hydraulic systems, using sensors. The operating data includes vibration amplitude, temperature changes, current fluctuations, and time-series change curves. The time-series running data is processed into frames according to a preset time window, and multiple sets of frame data are obtained by setting a preset frame overlap rate threshold. Extract the time-domain statistical features and frequency-domain features of each set of frame data. The time-domain statistical features include mean, variance, and peak factor. The frequency-domain features include the main frequency amplitude and harmonic components. The time-domain statistical features and frequency-domain features of each set of frame data are combined into a feature vector, and then arranged in frame order to construct a device operation feature matrix.

3. The artificial intelligence-based industrial equipment condition detection method according to claim 2, characterized in that, Based on lightweight convolutional kernels, hierarchical feature extraction is performed on the device operation feature matrix to obtain shallow time-domain features and deep frequency-domain features. The shallow time-domain features and deep frequency-domain features are then fused to obtain a fused feature vector, including: Construct a first lightweight convolutional kernel and a second lightweight convolutional kernel. The first lightweight convolutional kernel is used to extract local features in the temporal domain, and the second lightweight convolutional kernel is used to extract correlation features in the frequency domain. Based on the first lightweight convolutional kernel, shallow temporal features are obtained by calculating the device operation feature matrix. Based on the second lightweight convolution kernel, the shallow time-domain features are calculated to obtain the deep frequency-domain features; Dynamic weights are assigned to shallow time-domain features and deep frequency-domain features to obtain a fused feature vector.

4. The artificial intelligence-based industrial equipment condition detection method according to claim 3, characterized in that, Extract a fault feature template library of similar equipment from a historical equipment fault database, and calculate the similarity matching value between the fused feature vector and each template in the fault feature template library, including: Filter fault records that match the target equipment model and operating conditions from the historical equipment fault database, and extract the fault feature vectors corresponding to each fault record as fault feature templates; All fault feature templates are categorized and organized according to fault type to build a fault feature template library. Fault types include mechanical wear, electrical faults, and hydraulic leakage. Calculate the similarity between the fused feature vector and each template in the fault feature template library to obtain multiple similarity matching values; The system filters multiple similarity matching values, marking those with similarity matching values ​​less than or equal to a preset threshold as invalid matching results, and marking those with similarity matching values ​​greater than the preset threshold as valid matching results.

5. The artificial intelligence-based industrial equipment condition detection method according to claim 4, characterized in that, The probability of an anomaly in the device's operating status is determined based on the similarity matching value. If the anomaly probability is greater than a preset probability threshold, key operating parameters associated with the device anomaly are located based on the fused feature vector, including: The anomaly probability is calculated from the similarity matching values ​​in the valid matching results; If the probability of an anomaly is less than the preset probability threshold, the device is determined not to have any operational anomalies. If the probability of abnormality is greater than or equal to the preset probability threshold, the device is determined to have an operational abnormality. The contributing features are obtained by selecting the feature dimension that contributes the most to the anomaly probability from the fused feature vector; The contribution features are mapped to the corresponding device operating parameters, which are then identified as the key operating parameters associated with the anomalies.

6. The artificial intelligence-based industrial equipment condition detection method according to claim 5, characterized in that, Based on key operating parameters, a preliminary adjustment plan for equipment status is generated. The execution priority of this plan is dynamically optimized in conjunction with the equipment operating load, and a target maintenance plan is output, including: For each key operating parameter, a corresponding pre-selected adjustment plan is generated based on historical optimization data. The pre-selected adjustment plan includes the parameter adjustment range and adjustment steps. Collect current operating load data of the equipment, including load rate, continuous running time, and urgency of production tasks; Based on the operational load data and the impact range of the pre-selected adjustment schemes, the priority score of each pre-selected adjustment scheme is calculated; the pre-selected adjustment schemes are sorted from high to low according to their priority scores, integrated to form the target maintenance scheme, and the target maintenance scheme is output.

7. The artificial intelligence-based industrial equipment condition detection method according to claim 5, characterized in that, Also includes: The effectiveness of the target maintenance plan is monitored in real time, equipment operation data after the plan is implemented is collected, and new fusion feature vectors are extracted. Calculate the similarity matching value between the new fused feature vector and the fault feature template library to determine the probability of anomalies after the solution is implemented; If the probability of anomalies after the implementation of the solution is less than the preset probability threshold, the maintenance is deemed effective; if the probability of anomalies is greater than or equal to the preset probability threshold, the fault feature template library is updated and a pre-selected adjustment solution for the equipment status is regenerated.

Images