Industrial equipment lubricating oil life prediction method and system based on machine learning

Abstract

The invention provides an industrial equipment lubricating oil life prediction method and system based on machine learning, and the method comprises the steps: firstly obtaining a lubricating oil state monitoring data set of industrial equipment in a preset monitoring period, and enabling the lubricating oil state monitoring data set to comprise a plurality of lubricating state monitoring sequences; each lubricating state monitoring sequence is composed of lubricating oil attribute index data and equipment operation state data, then dynamic feature extraction is carried out on the lubricating oil state monitoring data set to obtain lubricating oil attribute dynamic change features and equipment operation state associated features, and then the lubricating oil state monitoring sequence is predicted based on a preset lubricating life prediction model. And finally, generating a lubrication maintenance decision instruction according to the lubrication oil life prediction value and a preset lubrication life threshold value, and feeding back the lubrication maintenance decision instruction to an industrial equipment maintenance terminal to trigger lubrication maintenance operation, thereby realizing effective prediction and maintenance decision of the service life of the industrial equipment lubrication oil.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular, to a method and system for predicting the lubricating oil life of industrial equipment based on machine learning. Background Art

[0002] In the field of industrial equipment operation and maintenance management, lubricating oil, as a key medium to ensure the normal operation of equipment, its performance state directly affects the reliability, efficiency and service life of the equipment. Traditional lubricating oil maintenance strategies mainly rely on regular replacement or passive maintenance modes based on simple thresholds (such as total running time, mileage or fixed cycle). Such methods have significant limitations: First, they do not fully consider the complexity of the dynamic change of lubricating oil performance with the operating conditions of the equipment, resulting in over-maintenance (causing waste of resources) or under-maintenance (leading to equipment failures); Second, existing monitoring technologies mostly focus on the off-line detection of single parameters (such as viscosity, acid value), lacking a comprehensive analysis of the correlation between multi-dimensional attributes of lubricating oil (such as oxidation degree, metal wear particle content, water content, etc.) and the operating state of the equipment (such as temperature, load, speed, etc.); Third, traditional methods are difficult to capture the non-linear characteristics of the deterioration of lubricating oil performance, unable to give early warnings of potential failure risks, resulting in maintenance decisions lagging behind actual needs.

[0003] In the prior art, some solutions attempt to monitor the lubricating oil state in real time through sensors, but only stay at the data acquisition level, without realizing the effective fusion and in-depth mining of multi-source heterogeneous data. For example, some systems only record the changes in the physical properties of lubricating oil, without correlating the real-time operating condition data of the equipment, resulting in one-sided analysis results; Another solution uses simple statistical models or threshold comparison for life prediction, but due to the single data dimension and insufficient algorithm robustness, the prediction accuracy is difficult to meet the requirements of industrial applications. In addition, existing maintenance decision-making mechanisms are mostly based on static rules, lacking dynamic adaptability, unable to adjust maintenance strategies according to the actual operating state of the equipment, resulting in low maintenance efficiency and high costs. Summary of the Invention

[0004] In view of the problems mentioned above, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for predicting the lubricating oil life of industrial equipment based on machine learning, the method comprising: Obtaining a set of lubricating oil state monitoring data of an industrial equipment within a preset monitoring period, the set of lubricating oil state monitoring data includes a plurality of lubricating state monitoring sequences, and each lubricating state monitoring sequence is composed of lubricating oil property index data and corresponding equipment operating state data at at least one lubricating state sampling moment; Performing dynamic feature extraction processing on the set of lubricating oil state monitoring data to obtain the dynamic change features of the lubricating oil properties and the equipment operating state correlation features of each lubricating state monitoring sequence; Based on a preset lubrication life prediction model, the dynamic change characteristics of the lubricating oil properties and the associated characteristics of the equipment operation status are fused and predicted to generate a lubricating oil life prediction value of the lubrication status monitoring sequence; A lubrication maintenance decision instruction is generated according to the lubrication oil life prediction value and a preset lubrication life threshold, and the lubrication maintenance decision instruction is fed back to the industrial equipment maintenance terminal to trigger a lubrication maintenance operation.

[0005] On the other hand, an embodiment of the present invention also provides an industrial equipment lubricant oil life prediction system based on machine learning, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0006] Based on the above aspects, the embodiment of the present application constructs a multi-dimensional time series data set including lubricating oil property indicators and equipment operating status. The dynamic feature extraction link analyzes the nonlinear law of the evolution of lubricating oil properties over time and the correlation characteristics with the equipment operating status, which not only captures the gradual process of lubrication performance degradation, but also reveals the synergistic influence mechanism of equipment load, temperature and other working parameters on lubrication life, thereby significantly improving the completeness and prediction accuracy of feature characterization. Multi-feature fusion prediction based on the preset lubrication life prediction model effectively overcomes the lag and one-sidedness of the traditional threshold judgment method, and deeply mines the potential patterns in the time series data through machine learning algorithms to achieve accurate quantitative evaluation of the remaining life of the lubricating oil. Finally, intelligent lubrication maintenance decision instructions are generated through dynamic threshold comparison, which not only reduces the risk of equipment failure caused by lubrication failure, but also optimizes lubrication resource consumption through accurate maintenance timing selection, significantly improving the reliability of industrial equipment operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a schematic diagram of the execution flow of the method for predicting the life of lubricating oil of industrial equipment based on machine learning provided in an embodiment of the present invention.

[0008] Figure 2 Schematic diagram of exemplary hardware and software components of an industrial equipment lubricant oil life prediction system based on machine learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1It is a schematic flowchart of a method for predicting the lubricating oil life of industrial equipment based on machine learning provided by an embodiment of the present invention. The method for predicting the lubricating oil life of industrial equipment based on machine learning will be introduced in detail below.

[0010] Step S110: Obtain a lubricating oil state monitoring data set of an industrial equipment within a preset monitoring period. The lubricating oil state monitoring data set includes a plurality of lubricating state monitoring sequences, and each lubricating state monitoring sequence is composed of lubricating oil property index data and corresponding equipment operation state data at at least one lubricating state sampling moment.

[0011] In this embodiment, taking a key processing equipment in a large manufacturing factory as an example, the preset monitoring period is set to 60 days. During these 60 days, a comprehensive sampling and monitoring of the lubricating oil is carried out every 3 days. Each time of sampling obtains a variety of lubricating oil property index data, such as lubricating oil viscosity, acid value, moisture content, oxidation stability, etc. At the same time, the equipment operation state data is recorded, such as equipment load, working temperature, vibration amplitude, rotation speed, etc. For example, when sampling on the 3rd day, the lubricating oil viscosity is 45 (unit: centistokes), the acid value is 0.4 (unit: mgKOH / g), the moisture content is 0.1%, and the oxidation stability test time is 800 (unit: minutes); the equipment load is 480 (unit: kilowatts), the working temperature is 58 (unit: degrees Celsius), the vibration amplitude is 0.05 (unit: millimeters), and the rotation speed is 1500 (unit: revolutions per minute), etc. According to this rule, multiple such sampling data groups can be formed within 60 days. The above sampling data groups constitute a plurality of lubricating state monitoring sequences, and then form a complete lubricating oil state monitoring data set.

[0012] Step S120: Perform dynamic feature extraction processing on the lubricating oil state monitoring data set to obtain the dynamic change characteristics of the lubricating oil properties and the equipment operation state correlation characteristics of each lubricating state monitoring sequence.

[0013] Next, the processing process of the lubricating oil state monitoring data set will be elaborated in detail.

[0014] Step S121: Perform time series segmentation processing on the lubricating oil property index data in the lubricating state monitoring sequence to obtain a plurality of lubricating property change segments.

[0015] In this embodiment, taking the lubricating oil viscosity data as an example, within the aforementioned 60-day monitoring period, a total of 20 sampling data are obtained. Then, these 20 sampling data can be segmented into segments with every 5 samplings as one segment. For example, assuming that the viscosity data of these 20 samplings are 45, 47, 50, 48, 49, 51, 53, 55, 54, 56, 58, 60, 59, 61, 63, 62, 64, 66, 65, 67 in sequence. Then, the data of the first segment are 45, 47, 50, 48, 49; the second segment is 51, 53, 55, 54, 56; the third segment is 58, 60, 59, 61, 63; the fourth segment is 62, 64, 66, 65, 67. Through such time series segmentation, multiple lubrication property change segments are obtained.

[0016] Step S122: Invoke the pre-set lubrication feature encoder to perform dynamic trend encoding processing on the multiple lubrication property change segments, and generate the dynamic change features of the lubricating oil properties of the lubrication state monitoring sequence, where the dynamic change features of the lubricating oil properties include the lubricating oil viscosity change rate, the acid value accumulation gradient, and the impurity concentration fluctuation trend.

[0017] The following further describes the sub-steps of this step: Step S1221: For each lubrication property change segment, extract the initial property index sequence in the lubrication property change segment and the corresponding monitoring time interval data.

[0018] For example, for the first viscosity change segment 45, 47, 50, 48, 49, the initial property index sequence is these 5 viscosity values, and the corresponding monitoring time interval data are all 3 days. For the acid value data, assuming that within the same sampling period, the first acid value change segment is 0.4, 0.42, 0.45, 0.43, 0.44, the initial property index sequence is these 5 acid values, and the monitoring time interval is also 3 days.

[0019] Step S1222: Perform local fluctuation analysis on the initial property index sequence based on a sliding time window to generate multiple local fluctuation feature subsequences.

[0020] For example, set the size of the sliding time window to 3. For the first viscosity change segment mentioned above, the first local fluctuation feature subsequence is 45, 47, 50. Analyze the fluctuation situation by calculating the difference or ratio of adjacent data. For example, calculate the differences between adjacent data as 47 - 45 = 2 and 50 - 47 = 3 respectively; the second local fluctuation feature subsequence is 47, 50, 48, and the calculated difference is 50 - 47 = 3 and 48 - 50 = -2; the third local fluctuation feature subsequence is 50, 48, 49, and the differences are 48 - 50 = -2 and 49 - 48 = 1. And so on, multiple local fluctuation feature subsequences are obtained. For the acid value change segment, it is analyzed in the same way.

[0021] Step S1223: Invoke the convolutional neural network layer in the lubrication feature encoder to perform convolutional processing on the local fluctuation feature subsequence to obtain a primary lubrication feature vector.

[0022] For example, assume that the size of the convolutional kernel of the convolutional neural network layer is 2 and the stride is 1. For the first viscosity local fluctuation feature subsequence 45, 47, 50, perform convolutional calculation. For example, perform a certain calculation (such as weighted summation, assuming the weights are 0.6 and 0.4 respectively) on the first convolutional kernel and the first two data 45, 47 to get 45×0.6 + 47×0.4 = 45.8; then slide the convolutional kernel to the next set of data 47, 50 and perform the same calculation to get 47×0.6 + 50×0.4 = 48.2. After a series of the above calculations, a primary lubrication feature vector is obtained, and this primary lubrication feature vector contains the feature information of the local fluctuation of this segment. The same operation is also performed on the local fluctuation feature subsequences of acid value and other attributes.

[0023] Step S1224: Obtain a weight adjustment coefficient corresponding to the monitoring time interval data, where the weight adjustment coefficient has a non-linear inverse proportional relationship with the duration of the monitoring time interval data.

[0024] For example, the weight adjustment coefficient for a monitoring time interval of 3 days is calculated through a certain non-linear function. Assume the non-linear function is f(x) = 1 / (x^2 + 2). When x = 3, the calculated weight adjustment coefficient is 1 / (3^2 + 2) = 1 / 11 ≈ 0.09. For different monitoring time intervals, the corresponding weight adjustment coefficients are calculated according to this non-linear function.

[0025] Step S1225: Multiply the weight adjustment coefficient element-wise with the primary lubrication feature vector to obtain a time-weighted lubrication feature vector.

[0026] In this embodiment, each weight adjustment coefficient can be multiplied by the element at the corresponding position in the primary lubrication feature vector. For example, if the primary lubrication feature vector is [45.8, 48.2,...] and the weight adjustment coefficient is 0.09, then the first element of the time-weighted lubrication feature vector is 45.8 × 0.09 = 4.122, and the second element is 48.2 × 0.09 = 4.338, and so on, to obtain the time-weighted lubrication feature vector.

[0027] Step S1226: Invoke the bidirectional long short-term memory network layer in the lubrication feature encoder to perform temporal dependence modeling on the time-weighted lubrication feature vector, and generate the dynamic trend encoding result of the lubrication attribute change segment.

[0028] The sub-steps of this step are further described: Step S12261: Input the time-weighted lubrication feature vector into the forward processing branch of the bidirectional long short-term memory network layer, and perform forward sequence processing on the time-weighted lubrication feature vector in chronological order to generate a forward hidden state sequence.

[0029] For example, at the first time step, the first element 4.122 of the time-weighted lubrication feature vector is input, and through internal network calculations (such as multiplying with a weight matrix, adding a bias term, passing through an activation function, etc.), the first forward hidden state is obtained; at the second time step, the second element 4.338 is input and combined with the first forward hidden state for similar calculations to obtain the second forward hidden state, and so on, to generate a forward hidden state sequence.

[0030] Step S12262: Synchronously input the time-weighted lubrication feature vector into the backward processing branch of the bidirectional long short-term memory network layer, and perform backward sequence processing on the time-weighted lubrication feature vector in reverse chronological order to generate a backward hidden state sequence.

[0031] That is, starting from the last time step, perform calculations similar to the forward processing in reverse. For example, at the last time step, the last element of the time-weighted lubrication feature vector is input, and through the calculation of the bidirectional long short-term memory network, the first backward hidden state is obtained; at the penultimate time step, the penultimate element is input and combined with the first backward hidden state for calculation to obtain the second backward hidden state, and thus a backward hidden state sequence is generated.

[0032] Step S12263: Perform cross-sequence feature alignment on the forward hidden state sequence and the backward hidden state sequence to obtain the concatenated feature vector of the forward hidden state and the backward hidden state at each time step.

[0033] For example, the forward hidden state and backward hidden state at the first time step are concatenated in the dimension. Assuming the forward hidden state is a 10-dimensional vector [a1, a2, …, a10] and the backward hidden state is a 10-dimensional vector [b1, b2, …, b10], the concatenated feature vector is [a1, a2, …, a10, b1, b2, …, b10], forming a new concatenated feature vector.

[0034] Step S12264: Perform non-linear activation and feature dimensionality reduction on the concatenated feature vector to generate the temporal dependence feature of the time-weighted lubrication feature vector at the corresponding time step.

[0035] The concatenated feature vector is processed through a non-linear activation function, such as the ReLU function (i.e., taking the larger value between the input value and 0). For example, for each element in the concatenated feature vector, if the element value is less than 0, it is set to 0. Then, feature dimensionality reduction is performed. For example, a fully connected layer is used to reduce the concatenated feature vector from a high dimension to a low dimension. Assuming it is reduced from 20 dimensions to 10 dimensions, the temporal dependence feature of each time step is obtained through calculation.

[0036] Step S12265: Aggregate the dynamic trend encoding results of all lubrication property change segments to generate the dynamic change feature of the lubricating oil properties of the lubrication state monitoring sequence.

[0037] In this embodiment, the dynamic trend encoding results obtained from all segments can be concatenated in the dimension to form a dynamic change feature of the lubricating oil properties including information such as the change rate of lubricating oil viscosity, the cumulative gradient of acid value, and the fluctuation trend of impurity concentration. For example, the dynamic trend encoding result of the first segment is a 10-dimensional vector, and the second segment is also a 10-dimensional vector. They are concatenated together to form a 20-dimensional vector, and so on, finally obtaining the complete dynamic change feature of the lubricating oil properties.

[0038] Step S123: Perform a working condition mode matching process on the equipment operation state data in the lubrication state monitoring sequence to obtain equipment operation state association features, where the equipment operation state association features include at least one of the following: equipment load fluctuation feature, temperature association influence coefficient, mechanical vibration frequency association degree.

[0039] The following further describes the sub-steps of this step: Step S1231: Extract the load fluctuation sequence, temperature change sequence, and vibration frequency sequence from the equipment operation state data.

[0040] For example, the load fluctuation sequence is the device load values recorded at each sampling within 60 days, assumed to be 480, 490, 500, 495, 485, 510, 520, 515, 505, 530, 540, 535, 525, 550, 560, 555, 545, 570, 565, 575 in sequence; the temperature change sequence is the temperature values recorded at each sampling, such as 58, 60, 63, 61, 59, 62, 64, 63, 62, 65, 66, 65, 64, 67, 68, 67, 66, 69, 68, 70; the vibration frequency sequence is the mechanical vibration frequency values recorded at each sampling, for example 50, 52, 55, 53, 54, 56, 58, 60, 59, 61, 63, 62, 64, 65, 66, 65, 64, 67, 66, 68.

[0041] Step S1232: Perform peak detection and valley alignment processing on the load fluctuation sequence to generate device load fluctuation characteristics.

[0042] In this embodiment, the peaks and valleys in the load fluctuation sequence can be detected through an algorithm with set rules. For example, after calculation, the peaks are 575, 570, 565, 560, 555, and the valleys are 480, 485, 490, 495, 500. Then, alignment processing is performed on the peaks and valleys, and they are sorted into a specific feature representation form according to certain rules to generate device load fluctuation characteristics. For example, the peaks and valleys can be combined into a new vector, or characteristic values representing the load fluctuation situation, such as the difference or average value between the peak and valley, can be calculated through a certain statistical method to form device load fluctuation characteristics.

[0043] Step S1233: Match the temperature change sequence with a preset temperature influence model to calculate the temperature correlation influence coefficient, where the temperature correlation influence coefficient is used to characterize the cumulative influence degree of temperature change on the lubricating oil property index.

[0044] For example, the sub-steps of step S1233 are as follows: Step S12331: Extract the historical temperature values and corresponding monitoring timestamps in the temperature change sequence.

[0045] For example, for the temperature change sequence 58, 60, 63, 61, 59, 62, 64, 63, 62, 65, 66, 65, 64, 67, 68, 67, 66, 69, 68, 70, the corresponding monitoring timestamps start from the 3rd day and are 3, 6, 9, 12, 15, 18, 21, 24, 27, 30, 33, 36, 39, 42, 45, 48, 51, 54, 57, 60 in sequence.

[0046] Step S12332: According to the reference temperature threshold and time window segmentation rule defined in the preset temperature impact model, segment the temperature change sequence into multiple temperature fluctuation time windows.

[0047] For example, assume the reference temperature threshold is 60 degrees Celsius and the time window segmentation rule is one window per 10 days. Then, the data included in the first temperature fluctuation time window are 58, 60, 63, 61, and the time range is from the 3rd day to the 12th day; the second window includes 59, 62, 64, 63, and the time range is from the 15th day to the 24th day; and so on, segmenting the entire temperature change sequence into multiple temperature fluctuation time windows.

[0048] Step S12333: For each temperature fluctuation time window, calculate the cumulative temperature deviation value between all historical temperature values within the window and the reference temperature threshold.

[0049] For example, for the first temperature fluctuation time window, the cumulative temperature deviation value is (58 - 60) + (60 - 60) + (63 - 60) + (61 - 60) = -2 + 0 + 3 + 1 = 2. For other windows, calculate the cumulative temperature deviation value in the same way.

[0050] Step S12334: Input the cumulative temperature deviation value corresponding to each temperature fluctuation time window into the preset non - linear impact function for mapping to obtain the window impact factor for each temperature fluctuation time window.

[0051] For example, assume the preset non - linear impact function is f(x)=x^2 + 1. For the cumulative temperature deviation value of 2 for the first window, the window impact factor obtained after mapping is 2^2 + 1 = 5. Perform this mapping operation on the cumulative temperature deviation value of each window to obtain the window impact factors of each window.

[0052] Step S12335: According to the time decay coefficient configured in the temperature impact model, perform time weighting on the window impact factor of each temperature fluctuation time window to generate a weighted window impact factor sequence.

[0053] For example, assume the time decay coefficient is 0.9, and the window impact factor of the first window is 5. The time weight is calculated according to the ratio of the start time of the window to the total monitoring period. For example, the start time of the first window is the 3rd day and the total monitoring period is 60 days, then the time weight is 3 / 60 = 0.05. The weighted window impact factor is 5×0.9^0.05≈4.9. Calculate the time weighting of the window impact factor of each window in this way to generate a weighted window impact factor sequence.

[0054] Step S12336: Accumulate and sum the weighted window influence factor sequence to generate the temperature correlation influence coefficient of the temperature change sequence.

[0055] In this embodiment, all the values in the weighted window influence factor sequence can be added. For example, if the weighted window influence factor sequence is [4.9, 5.2,...], the accumulated sum is used to obtain the temperature correlation influence coefficient. Suppose the final result is 30.

[0056] Step S1234: Determine the mechanical vibration frequency correlation degree according to the frequency domain energy distribution of the vibration frequency sequence. The mechanical vibration frequency correlation degree is used to quantify the accelerating attenuation effect of mechanical vibration on the lubricating oil life.

[0057] First, perform a frequency domain transformation on the vibration frequency sequence. For example, use the fast Fourier transform (FFT) to convert the vibration frequency sequence in the time domain into a frequency domain representation. Suppose the vibration frequency sequence is 50, 52, 55, 53, 54, 56, 58, 60, 59, 61, 63, 62, 64, 65, 66, 65, 64, 67, 66, 68. After the FFT transformation, the frequency domain energy distribution is obtained. Then, by analyzing certain characteristics of the frequency domain energy distribution, such as the energy proportion within a specific frequency range, the position of the energy peak, etc., to determine the mechanical vibration frequency correlation degree. For example, if the energy proportion within a certain key frequency range is relatively large, it indicates that the vibration at this frequency has a greater impact on the lubricating oil life. According to the pre-set rules, this correlation degree is quantified as a numerical value. Suppose it is 0.6, indicating that the mechanical vibration has a certain degree of accelerating attenuation effect on the lubricating oil life.

[0058] Step S130: Based on the pre-set lubricating life prediction model, fuse and predict the dynamic change characteristics of the lubricating oil properties and the correlation characteristics of the equipment operating state to generate the lubricating oil life prediction value of the lubricating state monitoring sequence.

[0059] The following further describes the sub-steps of this step: Step S131: Divide the dynamic change characteristics of the lubricating oil properties into a viscosity change characteristic subset, an acid value accumulation characteristic subset, and an impurity fluctuation characteristic subset.

[0060] For example, assume that the dynamic change characteristics of the lubricating oil fluid properties are a 30-dimensional vector. Assume that the first 10-dimensional data mainly reflects the change information of the lubricating oil viscosity. These 10-dimensional data can be extracted to form a viscosity change characteristic subset; the middle 10-dimensional data is mainly related to the acid value accumulation situation, and after extraction, it forms an acid value accumulation characteristic subset; the last 10-dimensional data reflects the fluctuation trend of the impurity concentration, constituting an impurity fluctuation characteristic subset. In this way, through the dimensionality division of the dynamic change characteristics of the lubricating oil fluid properties, the characteristics in different aspects are separated, facilitating subsequent targeted fusion processing with the characteristics related to the equipment operating state.

[0061] Step S132: Divide the characteristics related to the equipment operating state into a load-related characteristic subset, a temperature-related characteristic subset, and a vibration-related characteristic subset.

[0062] For the characteristics related to the equipment operating state, assume it is a 25-dimensional vector. Among them, the first 8-dimensional data mainly describes the fluctuation of the equipment load, and it is extracted as the load-related characteristic subset; the middle 8-dimensional data is related to the temperature-related influence coefficient and the comprehensive influence of temperature change on the equipment and the lubricating oil fluid, forming the temperature-related characteristic subset; the last 9-dimensional data reflects the mechanical vibration frequency correlation and other characteristics related to vibration, forming the vibration-related characteristic subset. This division enables different key factors of the equipment operating state to be effectively matched and fused with the lubricating oil fluid property characteristics respectively.

[0063] Step S133: Invoke the feature cross layer in the lubrication life prediction model to perform cross combination on the viscosity change characteristic subset and the load-related characteristic subset to generate a first cross feature vector.

[0064] The following further describes the sub-steps of this step: Step S1331: Extract the viscosity gradient sequence in the viscosity change characteristic subset and the load fluctuation amplitude sequence in the load-related characteristic subset.

[0065] In the viscosity change characteristic subset, the viscosity gradient sequence is obtained by calculating the difference between adjacent viscosity values. For example, if the 10-dimensional data corresponding to the viscosity change characteristic subset is [v1, v2, v3, v4, v5, v6, v7, v8, v9, v10], the calculated viscosity gradient sequence is [v2 - v1, v3 - v2, v4 - v3, v5 - v4, v6 - v5, v7 - v6, v8 - v7, v9 - v8, v10 - v9]. For the load-related characteristic subset, assume the 8-dimensional data is [l1, l2, l3, l4, l5, l6, l7, l8], and the load fluctuation amplitude sequence is obtained by calculating the absolute value of the difference between adjacent load values, such as [|l2 - l1|, |l3 - l2|, |l4 - l3|, |l5 - l4|, |l6 - l5|, |l7 - l6|, |l8 - l7|].

[0066] Step S1332: Calculate the dynamic correlation coefficient matrix between the viscosity gradient sequence and the load fluctuation amplitude sequence.

[0067] In this embodiment, a dynamic correlation analysis method can be used to consider the variation of the viscosity gradient sequence and the load fluctuation amplitude sequence over time. For example, calculate the correlation between the corresponding elements of the two sequences within different time windows. Assume the time window is 3. For the viscosity gradient sequence [g1, g2, g3, g4, g5] and the load fluctuation amplitude sequence [a1, a2, a3, a4, a5], calculate the correlation between g1 and a1, g2 and a2, g3 and a3 within the first time window, and so on, to obtain a dynamic correlation coefficient matrix, which can reflect the change in the degree of association between the viscosity gradient and the load fluctuation amplitude at different times.

[0068] Step S1333: Based on the dynamic correlation coefficient matrix, perform weighted correction on the viscosity gradient sequence to obtain a corrected viscosity gradient feature vector.

[0069] In this embodiment, the values of each element in the dynamic correlation coefficient matrix can be used as weights to perform weighted calculation on the viscosity gradient sequence. For example, if the elements at the corresponding positions in the dynamic correlation coefficient matrix are [w1, w2, w3, w4, w5], and the viscosity gradient sequence is [g1, g2, g3, g4, g5], then the first element of the corrected viscosity gradient feature vector is g1×w1, the second element is g2×w2, and so on, to obtain the corrected viscosity gradient feature vector. This weighted correction takes into account the dynamic relationship between the viscosity gradient and the load fluctuation amplitude, making the viscosity gradient feature more accurately reflect the information related to the load.

[0070] Step S1334: Concatenate the corrected viscosity gradient feature vector and the load fluctuation amplitude sequence to generate a first cross-feature vector.

[0071] In this embodiment, the corrected viscosity gradient feature vector and the load fluctuation amplitude sequence can be concatenated in terms of dimensions. Assume the corrected viscosity gradient feature vector is a 5-dimensional vector [m1, m2, m3, m4, m5], and the load fluctuation amplitude sequence is a 5-dimensional vector [a1, a2, a3, a4, a5], then the first cross-feature vector is [m1, m2, m3, m4, m5, a1, a2, a3, a4, a5]. Through this concatenation method, the associated features between the viscosity change and the load fluctuation are fused.

[0072] Step S134: Perform non-linear mapping processing on the acid value cumulative feature subset and the temperature correlation feature subset to generate a second cross-feature vector.

[0073] The sub - steps of this step are further described below: Step S1341: Obtain the acid value accumulation rate data in the acid value accumulation feature subset and the temperature influence coefficient in the temperature correlation feature subset.

[0074] In this embodiment, in the acid value accumulation feature subset, the acid value accumulation rate data is obtained by calculating the change in acid value within a certain time interval. For example, the 10 - dimensional data corresponding to the acid value accumulation feature subset reflects the acid value accumulation at different time points. Calculate the difference in acid value between adjacent time points divided by the time interval to obtain the acid value accumulation rate data. For the temperature correlation feature subset, directly extract the calculated temperature influence coefficient.

[0075] Step S1342: Input the acid value accumulation rate data into a preset exponential mapping function to obtain non - linear acid value accumulation features.

[0076] In this embodiment, assume that the preset exponential mapping function is f(x)=exp(x), and substitute each value in the acid value accumulation rate data into this function for calculation. For example, if the acid value accumulation rate data is [r1, r2, r3], the non - linear acid value accumulation features obtained after mapping are [exp(r1), exp(r2), exp(r3)]. This exponential mapping can highlight the change trend of the acid value accumulation rate and convert the linear acid value accumulation rate into a non - linear feature that can better reflect its impact on the lubricating oil life.

[0077] Step S1343: Multiply the non - linear acid value accumulation features and the temperature influence coefficient element - by - element to obtain a temperature - weighted acid value feature vector.

[0078] In this embodiment, assume that the non - linear acid value accumulation features are [n1, n2, n3] and the temperature influence coefficient is t. Then the first element of the temperature - weighted acid value feature vector is n1×t, the second element is n2×t, and the third element is n3×t. By this element - by - element multiplication method, the influence of temperature on acid value accumulation is incorporated into the feature vector, reflecting the potential influence of the interaction between temperature and acid value accumulation on the lubricating oil life.

[0079] Step S1344: Call the fully - connected layer in the lubricating life prediction model to perform dimensionality reduction processing on the temperature - weighted acid value feature vector to generate a second cross - feature vector.

[0080] In this embodiment, the fully connected layer performs matrix multiplication on the temperature-weighted acid value feature vector through a weight matrix, adds a bias term, and then according to the set dimensionality reduction target, for example, reduces the vector from 3D to 2D. Suppose the temperature-weighted acid value feature vector is [x1, x2, x3], the weight matrix is [[w11, w12], [w21, w22], [w31, w32]], and the bias term is [b1, b2]. After calculation, the second cross feature vector [y1, y2] after dimensionality reduction is obtained, where y1 = x1×w11 + x2×w21 + x3×w31 + b1, and y2 = x1×w12 + x2×w22 + x3×w32 + b2.

[0081] Step S135: Perform attention weight distribution processing on the impurity fluctuation feature subset and the vibration correlation feature subset to generate a third cross feature vector.

[0082] The following further describes the sub-steps of this step: Step S1351: Extract the impurity concentration fluctuation sequence in the impurity fluctuation feature subset and the vibration frequency correlation degree in the vibration correlation feature subset.

[0083] In this embodiment, in the impurity fluctuation feature subset, sequence data reflecting the fluctuation of impurity concentration over time is directly extracted. For example, the 10D data corresponding to the impurity fluctuation feature subset reflects the impurity concentration fluctuation at different sampling times, and this sequence is extracted as the impurity concentration fluctuation sequence. For the vibration correlation feature subset, the previously calculated mechanical vibration frequency correlation degree value is extracted.

[0084] Step S1352: Invoke the attention mechanism layer in the lubrication life prediction model to calculate the attention weight distribution between the impurity concentration fluctuation sequence and the vibration frequency correlation degree.

[0085] In this embodiment, the attention mechanism layer determines the attention weight distribution by calculating a certain similarity or correlation between the impurity concentration fluctuation sequence and the vibration frequency correlation degree. For example, using the dot product attention mechanism, each element in the impurity concentration fluctuation sequence is subjected to a dot product operation with the vibration frequency correlation degree, and then the above dot product result is normalized through the softmax function to obtain the attention weight distribution. Suppose the impurity concentration fluctuation sequence is [i1, i2, i3] and the vibration frequency correlation degree is v, then calculate the dot product [i1×v, i2×v, i3×v], and after being processed by the softmax function, the attention weight distribution [w1, w2, w3] is obtained, where w1 + w2 + w3 = 1, and each weight value represents the importance degree of this impurity concentration fluctuation element relative to the vibration frequency correlation degree.

[0086] Step S1353: Dynamically weight the impurity concentration fluctuation sequence according to the attention weight distribution to generate a weighted impurity fluctuation feature.

[0087] In this embodiment, the attention weight distribution and the impurity concentration fluctuation sequence can be multiplied element by element to obtain a weighted impurity fluctuation feature. For example, if the impurity concentration fluctuation sequence is [i1, i2, i3] and the attention weight distribution is [w1, w2, w3], then the weighted impurity fluctuation feature is [i1×w1, i2×w2, i3×w3]. This kind of dynamic weighting can highlight the more important part of the impurity concentration fluctuation in the case related to the vibration frequency correlation.

[0088] Step S1354: Fuse the weighted impurity fluctuation feature and the vibration frequency correlation to generate a third cross feature vector.

[0089] In this embodiment, the weighted impurity fluctuation feature and the vibration frequency correlation can be concatenated in dimension. Suppose the weighted impurity fluctuation feature is a 3D vector [j1, j2, j3] and the vibration frequency correlation is v, then the third cross feature vector is [j1, j2, j3, v]. Through this fusion method, the information of the impurity fluctuation and the vibration frequency correlation is integrated together.

[0090] Step S136: Fuse the first cross feature vector, the second cross feature vector and the third cross feature vector to generate a multi-dimensional lubricating oil feature set.

[0091] In this embodiment, the first cross feature vector, the second cross feature vector and the third cross feature vector can be concatenated in dimension. Suppose the first cross feature vector is a 10D vector [f1, f2, …, f10], the second cross feature vector is a 2D vector [s1, s2], and the third cross feature vector is a 4D vector [t1, t2, t3, t4], then the multi-dimensional lubricating oil feature set is [f1, f2, …, f10, s1, s2, t1, t2, t3, t4], forming a comprehensive feature set containing correlation features in multiple aspects such as viscosity and load, acid value and temperature, impurity and vibration.

[0092] Step S137: Call the regression prediction layer in the lubricating life prediction model to perform life prediction on the multi-dimensional lubricating oil feature set to generate the lubricating oil life prediction value.

[0093] In this embodiment, the regression prediction layer uses pre-trained parameters and a set regression algorithm, such as a linear regression or non-linear regression algorithm. Assuming a linear regression algorithm is used, the regression prediction layer includes a weight matrix W and a bias term b. The multi-dimensional lubricating oil feature set is an n-dimensional vector X = [x1, x2, …, xn], and the predicted lubricating oil life value y is obtained by calculating y = WX + b. The calculation process here involves matrix multiplication and addition operations. Each row of the weight matrix W corresponds to the weight of a feature dimension, and the influence degree of different features on the lubricating oil life is learned through pre-training. The bias term b is a constant term used to adjust the prediction result. For example, the weight matrix W is [[w11, w12, …, w1n], [w21, w22, …, w2n], …, [wm1, wm2, …, wmn]], the bias term b is [b1, b2, …, bm], and the calculated predicted value y is a scalar representing the predicted lubricating oil life.

[0094] Step S140: Generate a lubrication maintenance decision instruction based on the predicted lubricating oil life value and a preset lubricating life threshold, and feedback the lubrication maintenance decision instruction to the industrial equipment maintenance terminal to trigger a lubrication maintenance operation.

[0095] The following further describes the sub-steps of this step: Step S141: When the predicted lubricating oil life value is less than the first lubricating life threshold, generate a first lubrication maintenance decision instruction, which is used to indicate to immediately perform a lubricating oil replacement operation.

[0096] In this embodiment, assume that the first lubricating life threshold is set to 30 days, and the predicted lubricating oil life value calculated by the lubricating life prediction model is 20 days. Since 20 days is less than 30 days, a first lubrication maintenance decision instruction is generated at this time. This first lubrication maintenance decision instruction is sent to the relevant operation terminal through the industrial equipment maintenance system, clearly instructing the operator to immediately replace the lubricating oil of the industrial equipment to avoid equipment failure or damage caused by a serious decline in the performance of the lubricating oil.

[0097] Step S142: When the predicted lubricating oil life value is between the first lubricating life threshold and the second lubricating life threshold, generate a second lubrication maintenance decision instruction, which is used to trigger an operation to increase the real-time monitoring frequency of the lubricating oil state.

[0098] For example, the first lubrication life threshold is 30 days, the second lubrication life threshold is 45 days, and the predicted value of the lubricating oil life is 35 days. Since 30 days < 35 days < 45 days, a second lubrication maintenance decision instruction can be generated. This second lubrication maintenance decision instruction will cause the monitoring system of the industrial equipment to increase the real-time monitoring frequency of the lubricating oil state from once every 3 days to once every 1 day. By increasing the monitoring frequency, the state change of the lubricating oil can be grasped more timely, so as to take measures in time when the performance of the lubricating oil deteriorates significantly.

[0099] Step S143: When the predicted value of the lubricating oil life is greater than the second lubrication life threshold, generate a third lubrication maintenance decision instruction, and the third lubrication maintenance decision instruction is used to maintain the current lubrication maintenance strategy and output a remaining life prompt message.

[0100] In this embodiment, it is assumed that the second lubrication life threshold is 45 days and the predicted value of the lubricating oil life is 50 days. Since 50 days is greater than 45 days, a third lubrication maintenance decision instruction can be generated. This third lubrication maintenance decision instruction instructs the industrial equipment to continue operating according to the current lubrication maintenance strategy, and at the same time outputs a remaining life prompt message on the operation terminal or the relevant monitoring interface, such as prompting "The remaining life of the lubricating oil is about 50 days, and the current lubrication maintenance strategy can continue to be executed", so that the operator can understand the remaining available time of the lubricating oil and reasonably arrange the subsequent maintenance work.

[0101] In this embodiment, the lubrication maintenance decision instruction is sent to the industrial equipment maintenance terminal through the communication network of the industrial equipment. After receiving the lubrication maintenance decision instruction, the industrial equipment maintenance terminal can perform corresponding operations according to the instruction type. If it is the first lubrication maintenance decision instruction, the terminal will pop up a prompt box, clearly display "Immediately execute the lubricating oil replacement operation", and provide detailed operation steps and precautions to guide the operator to perform the replacement operation; if it is the second lubrication maintenance decision instruction, the terminal will automatically adjust the parameters of the monitoring system, increase the monitoring frequency of the lubricating oil state to the specified interval time, and record the relevant parameter adjustment information; if it is the third lubrication maintenance decision instruction, the terminal will display the remaining life prompt message on the main interface, while keeping the current lubrication maintenance strategy settings unchanged, ensuring that the lubrication maintenance work of the industrial equipment can be carried out according to a reasonable plan.

[0102] Based on the above steps, in the embodiments of this application, a multi-dimensional time series dataset including lubricating oil fluid property indicators and equipment operating states is constructed. In the dynamic feature extraction stage, by analyzing the non-linear law of the evolution of lubricating oil fluid properties over time and the correlation features with equipment operating states, not only the gradual process of lubrication performance degradation is captured, but also the collaborative influence mechanism of operating condition parameters such as equipment load and temperature on lubrication life is revealed, thus significantly improving the completeness of feature representation and prediction accuracy. Based on the pre-set lubrication life prediction model, multi-feature fusion prediction is carried out, effectively overcoming the lag and one-sidedness of traditional threshold judgment methods. By using machine learning algorithms to deeply mine the potential patterns in time series data, accurate quantitative evaluation of the remaining life of lubricating oil is achieved. Finally, intelligent lubrication maintenance decision instructions are generated through dynamic threshold comparison, not only reducing the equipment failure risk caused by lubrication failure, but also optimizing the consumption of lubrication resources by selecting accurate maintenance timing, significantly improving the reliability of industrial equipment operation.

[0103] Further, the training process of the lubrication life prediction model includes the following steps: Step S210: Obtain a sample monitoring dataset, where the sample monitoring dataset includes multiple sample monitoring sequences, and each sample monitoring sequence is composed of lubricating oil fluid property index data, equipment operating state data arranged in chronological order, and corresponding actual lubrication life annotation values.

[0104] In the actual process of collecting sample data, sample monitoring data can be obtained from multiple industrial equipment of the same type. For example, the operating data of 50 equipment within a period of time is collected. For each equipment, sampling monitoring is carried out at a certain time interval. Assuming the sampling interval is 2 days and the continuous monitoring is 90 days. Each time of sampling records various property index data of the lubricating oil, such as lubricating oil viscosity, acid value, moisture content, additive content, etc., and equipment operating state data, such as equipment load, working temperature, vibration frequency, rotation speed, etc. At the same time, through the actual equipment operation records and maintenance history, the actual lubrication life annotation value corresponding to each sampling period is determined. For example, for the first equipment, when sampling on the 2nd day, the lubricating oil viscosity is 42 (unit: centistokes), the acid value is 0.35 (unit: mgKOH / g), the moisture content is 0.08%, and the additive content is 3%; the equipment load is 450 (unit: kilowatt), the working temperature is 55 (unit: degree Celsius), the vibration frequency is 48 (unit: hertz), the rotation speed is 1450 (unit: revolutions per minute), and the actual lubrication life annotation value is 60 days. In this way, complete data is recorded for each sampling of each equipment within 90 days, forming multiple sample monitoring sequences, and all the above sequences constitute the sample monitoring dataset.

[0105] Step S220: Perform dynamic trend encoding processing on the lubricating oil property index data in each sample monitoring sequence to generate the dynamic change characteristics of the sample lubricating oil properties.

[0106] This step is similar to the dynamic trend encoding processing of the lubricating oil property index data in the previous step S122, and specifically may include the following steps Step S221: For each sample lubricating property change segment, extract the initial property index sequence and the corresponding monitoring time interval data in the lubricating property change segment.

[0107] For example, for the lubricating oil viscosity data in a certain sample monitoring sequence, within a specific time segment, the data is 42, 44, 46, 45, 47. The initial property index sequence is these 5 viscosity values, and the corresponding monitoring time interval data is 2 days for all.

[0108] Step S222: Perform local fluctuation analysis on the initial property index sequence based on a sliding time window to generate multiple local fluctuation characteristic subsequences.

[0109] For example, set the sliding time window size to 3. For the above viscosity sequence, the first local fluctuation characteristic subsequence is 42, 44, 46. Calculate the change rate of adjacent data, such as (44 - 42) / 42 ≈ 0.048, (46 - 44) / 44 ≈ 0.045, to characterize the fluctuation situation within this subsequence; the second local fluctuation characteristic subsequence is 44, 46, 45, and the calculated change rate is (46 - 44) / 44 ≈ 0.045, (45 - 46) / 46 ≈ -0.022; the third local fluctuation characteristic subsequence is 46, 45, 47, and the change rate is (45 - 46) / 46 ≈ -0.022, (47 - 45) / 45 ≈ 0.044. In this way, multiple local fluctuation characteristic subsequences are generated, and the above subsequences can reflect in detail the fluctuation characteristics of the lubricating oil property index in different local time periods.

[0110] Step S223: Call the convolutional neural network layer in the lubrication feature encoder to perform convolutional processing on the local fluctuation characteristic subsequences to obtain the primary lubrication feature vectors.

[0111] For example, assume that the convolution kernel size of the convolutional neural network layer is 2 and the stride is 1. For the first local fluctuation feature subsequence 42, 44, 46, the first convolution kernel is weighted with the first two data 42, 44 (assuming the weights are 0.7 and 0.3 respectively), resulting in 42×0.7 + 44×0.3 = 42.6; then the convolution kernel is slid to the next set of data 44, 46 and the same calculation is performed, obtaining 44×0.7 + 46×0.3 = 44.6. Such convolution calculations are performed on all local fluctuation feature subsequences, and finally a primary lubrication feature vector is obtained, which integrates the key information in the local fluctuation feature subsequences.

[0112] Step S224: Obtain a weight adjustment coefficient corresponding to the monitored time interval data, where the weight adjustment coefficient has a non-linear inverse proportional relationship with the duration of the monitored time interval data.

[0113] For example, the function f(x) = 1 / (x^2 + 3) can be used to calculate the weight adjustment coefficient. When the monitored time interval x = 2 days, the calculated weight adjustment coefficient is 1 / (2^2 + 3) = 1 / 7 ≈ 0.143. For different monitored time intervals in different sample monitoring sequences, the corresponding weight adjustment coefficients are calculated according to this function to reflect the influence of the time interval on the importance of the data.

[0114] Step S225: Multiply the weight adjustment coefficient element-wise with the primary lubrication feature vector to obtain a time-weighted lubrication feature vector.

[0115] For example, assume that the primary lubrication feature vector is [v1, v2, v3] and the weight adjustment coefficient is 0.143. Then the first element of the time-weighted lubrication feature vector is v1×0.143, the second element is v2×0.143, and the third element is v3×0.143. Through this element-wise multiplication operation, the influence of the time factor on the feature vector is incorporated, making the feature vector more accurately reflect the dynamic changes of the lubricating oil properties at different time scales.

[0116] Step S226: Invoke the bidirectional long short-term memory network layer in the lubrication feature encoder to perform temporal dependency modeling on the time-weighted lubrication feature vector, and generate a dynamic trend coding result of the lubrication attribute change segment.

[0117] Step S2261: Input the time-weighted lubrication feature vector into the forward processing branch of the bidirectional long short-term memory network layer, and perform forward sequence processing on the time-weighted lubrication feature vector in chronological order to generate a forward hidden state sequence.

[0118] For example, at the first time step, the first element of the time-weighted lubrication feature vector is input into the bidirectional long short-term memory network layer. After multiplying with the weight matrix inside the bidirectional long short-term memory network layer, adding the bias term, and processing through an activation function (such as the tanh function), the first forward hidden state is obtained. At the second time step, the second element and the first forward hidden state are input into the network for calculation to obtain the second forward hidden state, and so on, generating a sequence of forward hidden states. In this process, the bidirectional long short-term memory network layer learns the feature changes of the time-weighted lubrication feature vector in chronological order.

[0119] Step S2262: Synchronously input the time-weighted lubrication feature vector into the backward processing branch of the bidirectional long short-term memory network layer, perform backward sequence processing on the time-weighted lubrication feature vector in reverse chronological order, and generate a sequence of backward hidden states.

[0120] For example, starting from the last element of the time-weighted lubrication feature vector, it is input into the backward processing branch network. After a calculation process similar to the forward processing, the first backward hidden state is obtained. Then, the penultimate element and the first backward hidden state are input into the network to calculate the second backward hidden state, and so on to generate a sequence of backward hidden states. Through the backward processing, the network captures the feature dependency relationships in reverse chronological order.

[0121] Step S2263: Perform cross-sequence feature alignment on the sequence of forward hidden states and the sequence of backward hidden states to obtain the concatenated feature vector of the forward hidden state and the backward hidden state at each time step.

[0122] For example, concatenate the sequence of forward hidden states and the sequence of backward hidden states at each time step. For example, at the first time step, the forward hidden state is an 8-dimensional vector [h1_1, h1_2,..., h1_8], and the backward hidden state is an 8-dimensional vector [h2_1, h2_2,..., h2_8]. Then the concatenated feature vector is [h1_1, h1_2,..., h1_8, h2_1, h2_2,..., h2_8], forming a new vector that contains both forward and backward feature information.

[0123] Step S2264: Perform non-linear activation and feature dimensionality reduction processing on the concatenated feature vector to generate the temporal dependency features of the time-weighted lubrication feature vector at the corresponding time step.

[0124] For example, a non-linear activation function, such as the ReLU function, is used to process the concatenated feature vector, and elements less than 0 are set to 0. Then, feature dimensionality reduction is performed through a fully connected layer. For example, the 16-dimensional concatenated feature vector is reduced to 10 dimensions, and the temporal dependence features at each time step are obtained through calculation. The above features reflect the temporal dependence relationship of the time-weighted lubrication feature vector at different time steps.

[0125] Step S2265: Aggregate the dynamic trend encoding results of all lubrication property change segments to generate the dynamic change features of the sample lubricating oil properties.

[0126] In this embodiment, the dynamic trend encoding results obtained after the above processing of all lubrication property change segments are concatenated. For example, in a sample monitoring sequence, there are 5 lubrication property change segments, and the dynamic trend encoding result of each segment is a 10-dimensional vector. These 5 10-dimensional vectors are concatenated in sequence to form a 50-dimensional dynamic change feature vector of the sample lubricating oil properties, comprehensively reflecting the dynamic change of the lubricating oil properties in the sample.

[0127] Step S230: Perform a working condition mode matching process on the device operation state data in each sample monitoring sequence to generate sample device operation state association features.

[0128] This step is similar to the working condition mode matching process for the device operation state data in step S123 above.

[0129] Step S231: Extract the load fluctuation sequence, temperature change sequence, and vibration frequency sequence from the sample device operation state data.

[0130] For example, for a certain sample monitoring sequence, the device load fluctuation sequence is 450, 460, 470, 465, 455, etc. within 90 days; the temperature change sequence is 55, 57, 59, 58, 56, etc.; the vibration frequency sequence is 48, 50, 52, 51, 49, etc. The above sequences record the key parameters of the device operation state at different time points.

[0131] Step S232: Perform peak detection and valley alignment processing on the load fluctuation sequence to generate sample device load fluctuation features.

[0132] In this embodiment, the peaks and valleys in the load fluctuation sequence can be detected through a set algorithm. For example, after calculation, the peaks are found to be 470, 465, etc., and the valleys are 450, 455, etc. Then, the peaks and valleys are aligned, such as arranging them in the order of appearance, or calculating statistical quantities such as the difference or average value between the peaks and valleys to form sample device load fluctuation features for describing the fluctuation characteristics of the device load.

[0133] Step S233: Match the temperature change sequence with a preset temperature impact model, and calculate the sample temperature correlation impact coefficient.

[0134] Step S2331: Extract the historical temperature values and corresponding monitoring timestamps in the temperature change sequence.

[0135] Suppose the temperature change sequence is 55, 57, 59, 58, 56, and the corresponding monitoring timestamps are 2, 4, 6, 8, 10 days respectively.

[0136] Step S2332: According to the reference temperature threshold and time window segmentation rule defined in the preset temperature impact model, divide the temperature change sequence into multiple temperature fluctuation time windows.

[0137] Suppose the reference temperature threshold is 57 degrees Celsius, and the time window segmentation rule is one window every 4 days. Then, the first temperature fluctuation time window contains data 55, 57, and the time range is from the 2nd day to the 5th day; the second window contains 59, 58, and the time range is from the 6th day to the 9th day; the third window contains 56, and the time range is the 10th day.

[0138] Step S2333: For each temperature fluctuation time window, calculate the cumulative temperature deviation value between all historical temperature values in the window and the reference temperature threshold.

[0139] For the first temperature fluctuation time window, the cumulative temperature deviation value is (55 - 57) + (57 - 57) = -2. For the second window, the cumulative temperature deviation value is (59 - 57) + (58 - 57) = 3. For the third window, the cumulative temperature deviation value is 56 - 57 = -1.

[0140] Step S2334: Input the cumulative temperature deviation value corresponding to each temperature fluctuation time window into a preset non - linear impact function for mapping to obtain the window impact factor for each temperature fluctuation time window.

[0141] Suppose the preset non - linear impact function is f(x)=x^2 + 0.5. For the cumulative temperature deviation value -2 of the first window, the mapped window impact factor is (-2)^2 + 0.5 = 4.5. For the cumulative temperature deviation value 3 of the second window, the mapped window impact factor is 3^2 + 0.5 = 9.5. For the cumulative temperature deviation value -1 of the third window, the mapped window impact factor is (-1)^2 + 0.5 = 1.5.

[0142] Step S2335: According to the time decay coefficient configured in the temperature impact model, perform time weighting on the window impact factor of each temperature fluctuation time window to generate a weighted window impact factor sequence.

[0143] Assume that the time decay coefficient is 0.95, the window influence factor of the first window is 4.5, the time weight is calculated according to the ratio of the start time of the window to the total monitoring time. The start time of the first window is the 2nd day, and the total monitoring time is 90 days, so the time weight is 2 / 90. The weighted window influence factor is 4.5×0.95^(2 / 90)≈4.49. According to this method, the window influence factors of each window are weighted by time to generate a weighted window influence factor sequence.

[0144] Step S2336: Accumulate and sum the weighted window influence factor sequence to generate the sample temperature correlation influence coefficient.

[0145] Add all the values in the weighted window influence factor sequence. Assume the weighted window influence factor sequence is [4.49, 9.45, 1.48], and the accumulated sum gives the sample temperature correlation influence coefficient as 4.49 + 9.45 + 1.48 = 15.42.

[0146] Step S234: Determine the sample mechanical vibration frequency correlation degree according to the frequency domain energy distribution of the vibration frequency sequence.

[0147] First, perform a frequency domain transformation on the vibration frequency sequence. For example, use the Fast Fourier Transform (FFT) to convert the vibration frequency sequence in the time domain into a frequency domain representation. Assume the vibration frequency sequence is 48, 50, 52, 51, 49. After the FFT transformation, the frequency domain energy distribution is obtained. Analyze the frequency domain energy distribution. For example, it is found that the energy in a certain specific frequency range accounts for a relatively high proportion of the total energy. According to the pre-set rules, this proportion is used as the sample mechanical vibration frequency correlation degree. Assume that the energy proportion in a certain key frequency range is 0.6, then the sample mechanical vibration frequency correlation degree is 0.6.

[0148] Integrate the sample device load fluctuation characteristics, the sample temperature correlation influence coefficient, the sample mechanical vibration frequency correlation degree, etc. to form the sample device operation state correlation characteristics, which comprehensively reflect various factors related to the operation state of the device in the sample and the lubricating oil life.

[0149] Step S240: Input the sample lubricating oil property dynamic change characteristics and the sample device operation state correlation characteristics into the initial lubricating life prediction model, and perform multi-dimensional feature fusion through the feature cross layer to generate the sample fusion feature vector.

[0150] Input the dynamic change characteristics of the sample lubricating oil fluid properties and the correlation characteristics of the sample equipment operating status into the feature cross-layer of the initial lubrication life prediction model respectively. For example, the dynamic change characteristics of the sample lubricating oil fluid properties is a 50-dimensional vector, and the correlation characteristics of the sample equipment operating status is a 20-dimensional vector. The feature cross-layer can perform fusion operations on these two feature vectors in various ways.

[0151] For example, the feature cross-layer can calculate the correlation between different dimensions. For example, calculate the dot product of a certain dimension in the dynamic change characteristics of the lubricating oil fluid properties and a certain dimension in the correlation characteristics of the equipment operating status to obtain the correlation value. Then, perform a weighted combination of the feature vectors according to the above correlation value. For example, perform a weighted combination of the viscosity change-related dimension in the dynamic change characteristics of the lubricating oil fluid properties and the load fluctuation-related dimension in the correlation characteristics of the equipment operating status to generate a new part of the feature vector. Through such multi-dimensional feature cross and fusion operations, finally generate a sample fusion feature vector, which contains the correlation information between the lubricating oil fluid properties and the equipment operating status.

[0152] Step S250: Call the regression prediction layer in the initial lubrication life prediction model to perform life prediction on the sample fusion feature vector, and generate a sample lubrication life prediction value.

[0153] In this embodiment, the regression prediction layer uses the sample fusion feature vector for life prediction. Assume that the regression prediction layer adopts a linear regression model, and the linear regression model includes a weight matrix W and a bias term b. The sample fusion feature vector is an n-dimensional vector X = [x1, x2,..., xn], and the sample lubrication life prediction value y is obtained by calculating y = WX + b. Each row of the weight matrix W corresponds to the weight of a feature dimension, and the above weights are randomly initialized before model training. The bias term b is a constant vector. For example, the weight matrix W is [[w11, w12,..., w1n], [w21, w22,..., w2n],..., [wm1, wm2,..., wmn]], and the bias term b is [b1, b2,..., bm]. After matrix multiplication and addition operations, the prediction value y is obtained, and this value is the prediction result of the model for the sample lubricating oil life.

[0154] Step S260: Calculate the prediction loss based on the difference between the sample lubrication life prediction value and the actual lubrication life annotation value, and optimize the parameters of the initial lubrication life prediction model through backpropagation until the prediction loss converges, and generate a trained lubrication life prediction model.

[0155] In this embodiment, the prediction loss is usually calculated using a loss function, such as the mean squared error (MSE) loss function. Suppose the predicted value of the sample lubrication life is y_pred, and the actual labeled value of the lubrication life is y_true. Then the loss value L is the average value of (y_pred - y_true)^2. For example, for a sample monitoring sequence, the predicted value is 55 days, and the actual labeled value is 60 days. Then the loss value is (55 - 60)^2 = 25. The average of the loss values calculated for all sample monitoring sequences is taken to obtain the total prediction loss.

[0156] Then, the parameters of the initial lubrication life prediction model are optimized through the backpropagation algorithm. The backpropagation algorithm can calculate the gradients of each parameter in the lubrication life prediction model according to the prediction loss. For example, it calculates the gradients of the weight matrix W and the bias term b. According to the above gradients, an optimizer (such as Stochastic Gradient Descent SGD, Adagrad, Adam, etc.) is used to adjust the parameters. Taking Stochastic Gradient Descent as an example, the update formula for the weight matrix W is W = W - learning_rate * gradient(W), where learning_rate is the learning rate used to control the step size of parameter update, and gradient(W) is the gradient of the weight matrix W. The bias term b is updated in a similar manner.

[0157] By continuously repeating the above process of calculating the prediction loss and optimizing the parameters through backpropagation until the prediction loss converges. Convergence means that as the training progresses, the prediction loss no longer decreases significantly. At this time, the parameters of the model reach a relatively stable state, and a trained lubrication life prediction model is generated. This lubrication life prediction model can more accurately predict the life of the lubricating oil based on the dynamically changing characteristics of the input lubricating oil properties and the associated characteristics of the equipment operating state, providing a reliable decision-making basis for the lubrication maintenance of industrial equipment.

[0158] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of a machine learning-based industrial equipment lubricating oil life prediction system 100 that can implement the ideas of the present application provided in some embodiments of the present application. For example, the processor 120 can be used on the machine learning-based industrial equipment lubricating oil life prediction system 100 and is used to execute the functions in the present application.

[0159] The machine learning-based industrial equipment lubricating oil life prediction system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the machine learning-based industrial equipment lubricating oil life prediction method of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0160] For example, the machine learning-based industrial equipment lubricating oil life prediction system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROMs, or RAMs, or any combination thereof. Exemplarily, the machine learning-based industrial equipment lubricating oil life prediction system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to the above program instructions. The machine learning-based industrial equipment lubricating oil life prediction system 100 further includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0161] For ease of explanation, only one processor is described in the machine learning-based industrial equipment lubricating oil life prediction system 100. However, it should be noted that the machine learning-based industrial equipment lubricating oil life prediction system 100 in the present application may also include multiple processors. Therefore, the steps performed by one processor described in the present application can also be jointly executed or separately executed by multiple processors. For example, if the processor of the machine learning-based industrial equipment lubricating oil life prediction system 100 executes step A and step B, it should be understood that step A and step B can also be jointly executed by two different processors or separately executed in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor jointly execute steps A and B.

[0162] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned machine learning-based industrial equipment lubricating oil life prediction method is implemented.

[0163] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A method for predicting the lubricating oil life of industrial equipment based on machine learning, characterized in that, The method includes: Obtaining a lubricating oil liquid state monitoring data set of an industrial device within a preset monitoring period, where the lubricating oil liquid state monitoring data set includes multiple lubrication state monitoring sequences, and each lubrication state monitoring sequence is composed of lubricating oil liquid attribute index data and corresponding device operation state data at at least one lubrication state sampling moment; Performing dynamic feature extraction processing on the lubricating oil liquid state monitoring data set to obtain the dynamic change characteristics of the lubricating oil liquid attributes and the device operation state correlation characteristics of each lubrication state monitoring sequence; Based on a preset lubricating life prediction model, fusing and predicting the dynamic change characteristics of the lubricating oil liquid attributes and the device operation state correlation characteristics to generate a lubricating oil liquid life prediction value of the lubrication state monitoring sequence; Generating a lubrication maintenance decision instruction according to the lubricating oil liquid life prediction value and a preset lubricating life threshold, and feeding back the lubrication maintenance decision instruction to an industrial device maintenance terminal to trigger a lubrication maintenance operation.

2. The method for predicting the service life of lubricating oil for industrial equipment based on machine learning according to claim 1, wherein The performing dynamic feature extraction processing on the lubricating oil liquid state monitoring data set to obtain the dynamic change characteristics of the lubricating oil liquid attributes and the device operation state correlation characteristics of each lubrication state monitoring sequence includes: Performing time series segmentation processing on the lubricating oil liquid attribute index data in the lubrication state monitoring sequence to obtain multiple lubricating attribute change segments; Invoking a preset lubricating feature encoder to perform dynamic trend encoding processing on the multiple lubricating attribute change segments to generate the dynamic change characteristics of the lubricating oil liquid attributes of the lubrication state monitoring sequence, where the dynamic change characteristics of the lubricating oil liquid attributes include the lubricating oil viscosity change rate, the acid value accumulation gradient, and the impurity concentration fluctuation trend; Performing working condition mode matching processing on the device operation state data in the lubrication state monitoring sequence to obtain device operation state correlation characteristics, where the device operation state correlation characteristics include at least one of the following: device load fluctuation characteristics, temperature correlation influence coefficient, mechanical vibration frequency correlation degree.

3. The method for predicting the service life of lubricating oil for industrial equipment based on machine learning according to claim 2, wherein, The invoking a preset lubricating feature encoder to perform dynamic trend encoding processing on the multiple lubricating attribute change segments to generate the dynamic change characteristics of the lubricating oil liquid attributes of the lubrication state monitoring sequence includes: For each lubricating attribute change segment, extracting the initial attribute index sequence and the corresponding monitoring time interval data in the lubricating attribute change segment; Performing local fluctuation analysis on the initial attribute index sequence based on a sliding time window to generate multiple local fluctuation feature subsequences; Invoking a convolutional neural network layer in the lubricating feature encoder to perform convolution processing on the local fluctuation feature subsequences to obtain a primary lubricating feature vector; Obtaining a weight adjustment coefficient corresponding to the monitoring time interval data, where the weight adjustment coefficient has a non-linear inverse proportional relationship with the duration of the monitoring time interval data; Multiplying the weight adjustment coefficient and the primary lubricating feature vector element by element to obtain a time-weighted lubricating feature vector; Invoking a bidirectional long short-term memory network layer in the lubricating feature encoder to perform time series dependence modeling on the time-weighted lubricating feature vector to generate a dynamic trend encoding result of the lubricating attribute change segment; Aggregate the dynamic trend coding results of all lubrication property change segments to generate the dynamic change characteristics of the lubricating oil properties in the lubrication state monitoring sequence.

4. The method for predicting the lubricant life of industrial equipment based on machine learning according to claim 3, wherein The step of calling the bidirectional long short-term memory network layer in the lubrication feature encoder to perform temporal dependence modeling on the time-weighted lubrication feature vector and generate the dynamic trend coding result of the lubrication property change segment includes: Input the time-weighted lubrication feature vector into the forward processing branch of the bidirectional long short-term memory network layer, and perform forward sequence processing on the time-weighted lubrication feature vector in chronological order to generate a forward hidden state sequence; Synchronously input the time-weighted lubrication feature vector into the backward processing branch of the bidirectional long short-term memory network layer, and perform backward sequence processing on the time-weighted lubrication feature vector in reverse chronological order to generate a backward hidden state sequence; Perform cross-sequence feature alignment on the forward hidden state sequence and the backward hidden state sequence to obtain the concatenated feature vector of the forward hidden state and the backward hidden state at each time step; Perform non-linear activation and feature dimensionality reduction processing on the concatenated feature vector to generate the temporal dependence feature of the time-weighted lubrication feature vector at the corresponding time step; Aggregate the temporal dependence features of all time steps to generate the dynamic trend coding result of the lubrication property change segment.

5. The method for predicting the service life of lubricating oil for industrial equipment based on machine learning according to claim 2, wherein The step of performing a working condition mode matching process on the device operating state data in the lubrication state monitoring sequence to obtain the device operating state correlation feature includes: Extract the load fluctuation sequence, temperature change sequence, and vibration frequency sequence from the device operating state data; Perform peak detection and valley alignment processing on the load fluctuation sequence to generate the device load fluctuation feature; Match the temperature change sequence with a preset temperature influence model, and calculate the temperature correlation influence coefficient, where the temperature correlation influence coefficient is used to characterize the cumulative influence degree of temperature change on the lubricating oil property index; Determine the mechanical vibration frequency correlation degree according to the frequency domain energy distribution of the vibration frequency sequence, and the mechanical vibration frequency correlation degree is used to quantify the accelerating attenuation effect of mechanical vibration on the lubricating oil life.

6. The method for predicting the service life of lubricating oil for industrial equipment based on machine learning according to claim 1, characterized in that, The step of performing a fusion prediction on the dynamic change characteristics of the lubricating oil properties and the device operating state correlation feature based on a preset lubricating oil life prediction model to generate the lubricating oil life prediction value of the lubrication state monitoring sequence includes: Divide the dynamic change characteristics of the lubricating oil properties into a viscosity change characteristic subset, an acid value accumulation characteristic subset, and an impurity fluctuation characteristic subset; Divide the device operating state correlation feature into a load correlation feature subset, a temperature correlation feature subset, and a vibration correlation feature subset; Call the feature cross layer in the lubricating oil life prediction model to perform cross combination on the viscosity change characteristic subset and the load correlation feature subset to generate a first cross feature vector; Perform non-linear mapping processing on the acid value accumulation characteristic subset and the temperature correlation feature subset to generate a second cross feature vector; Perform attention weight assignment processing on the impurity fluctuation characteristic subset and the vibration correlation feature subset to generate a third cross feature vector; Fuse the first cross feature vector, the second cross feature vector, and the third cross feature vector to generate a multi-dimensional lubricating oil feature set; Call the regression prediction layer in the lubricating life prediction model to perform life prediction on the multi-dimensional lubricating oil feature set, and generate the lubricating oil life prediction value.

7. The method for predicting the service life of lubricating oil for industrial equipment based on machine learning according to claim 6, wherein The feature cross layer in the lubricating life prediction model is called to perform cross combination on the viscosity change feature subset and the load correlation feature subset to generate a first cross feature vector, including: Extract the viscosity gradient sequence in the viscosity change feature subset and the load fluctuation amplitude sequence in the load correlation feature subset; Calculate the dynamic correlation coefficient matrix between the viscosity gradient sequence and the load fluctuation amplitude sequence; Based on the dynamic correlation coefficient matrix, perform weighted correction on the viscosity gradient sequence to obtain a corrected viscosity gradient feature vector; Perform feature splicing on the corrected viscosity gradient feature vector and the load fluctuation amplitude sequence to generate a first cross feature vector; And, perform non-linear mapping processing on the acid value accumulation feature subset and the temperature correlation feature subset to generate a second cross feature vector, including: Obtain the acid value accumulation rate data in the acid value accumulation feature subset and the temperature influence coefficient in the temperature correlation feature subset; Input the acid value accumulation rate data into a preset exponential mapping function to obtain a non-linear acid value accumulation feature; Perform element-wise multiplication on the non-linear acid value accumulation feature and the temperature influence coefficient to obtain a temperature-weighted acid value feature vector; Call the fully connected layer in the lubricating life prediction model to perform dimensionality reduction processing on the temperature-weighted acid value feature vector to generate a second cross feature vector; And, perform attention weight distribution processing on the impurity fluctuation feature subset and the vibration correlation feature subset to generate a third cross feature vector, including: Extract the impurity concentration fluctuation sequence in the impurity fluctuation feature subset and the vibration frequency correlation degree in the vibration correlation feature subset; Call the attention mechanism layer in the lubricating life prediction model to calculate the attention weight distribution between the impurity concentration fluctuation sequence and the vibration frequency correlation degree; Perform dynamic weighting on the impurity concentration fluctuation sequence according to the attention weight distribution to generate a weighted impurity fluctuation feature; Perform feature fusion on the weighted impurity fluctuation feature and the vibration frequency correlation degree to generate a third cross feature vector.

8. The method for predicting the service life of lubricating oil for industrial equipment based on machine learning according to claim 1, characterized in that, Generating a lubrication maintenance decision instruction according to the lubricating oil life prediction value and a preset lubricating life threshold value, including: When the lubricating oil life prediction value is less than the first lubricating life threshold value, generate a first lubrication maintenance decision instruction, and the first lubrication maintenance decision instruction is used to indicate to immediately perform the lubricating oil replacement operation; When the lubricating oil life prediction value is between the first lubricating life threshold value and the second lubricating life threshold value, generate a second lubrication maintenance decision instruction, and the second lubrication maintenance decision instruction is used to trigger an operation to increase the real-time monitoring frequency of the lubricating oil state; When the predicted value of the lubricating oil life is greater than the second lubrication life threshold, a third lubrication maintenance decision instruction is generated, and the third lubrication maintenance decision instruction is used to maintain the current lubrication maintenance strategy and output a remaining life prompt message.

9. The method for predicting the service life of lubricating oil for industrial equipment based on machine learning according to claim 1, wherein The training process of the lubrication life prediction model includes: Obtain a sample monitoring data set, where the sample monitoring data set includes a plurality of sample monitoring sequences, and each sample monitoring sequence is composed of lubricating oil property index data, equipment operation status data arranged in chronological order, and corresponding actual lubrication life annotation values; Perform dynamic trend coding processing on the lubricating oil property index data in each sample monitoring sequence to generate sample lubricating oil property dynamic change characteristics; Perform working condition mode matching processing on the equipment operation status data in each sample monitoring sequence to generate sample equipment operation status correlation characteristics; Input the sample lubricating oil property dynamic change characteristics and the sample equipment operation status correlation characteristics into the initial lubrication life prediction model, and perform multi-dimensional feature fusion through the feature cross layer to generate a sample fusion feature vector; Call the regression prediction layer in the initial lubrication life prediction model to perform life prediction on the sample fusion feature vector to generate a sample lubrication life prediction value; Calculate the prediction loss based on the difference between the sample lubrication life prediction value and the actual lubrication life annotation value, and optimize the parameters of the initial lubrication life prediction model through backpropagation until the prediction loss converges, and generate a trained lubrication life prediction model.

10. An industrial equipment lubricating oil life prediction system based on machine learning, characterized in that, The machine learning-based industrial equipment lubricating oil life prediction system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the machine learning-based industrial equipment lubricating oil life prediction method according to any one of the above claims 1-9.

Images