Servo Drive System State Monitoring Method and System Based on Multi-Sensor Fusion

Abstract

The present invention provides a method and system for monitoring the state of a servo drive system based on multi-sensor fusion. First, a multi-source sensor data set including vibration waveforms, temperature gradients, and current harmonic data of a target servo drive device is obtained, and time synchronization and amplitude calibration are performed on it to generate a standardized monitoring data set. Then, a dynamic coding model is called to extract operation characteristics, forming an operation characteristic set. Next, the operation characteristic set is evaluated based on a multi-task state evaluation model to obtain real-time health state indicators, abnormal risk levels, and performance degradation trends. Finally, an equipment maintenance optimization strategy is generated based on the above results and fed back to the operation and maintenance control platform to trigger proactive maintenance operations. Thus, accurate state monitoring of the servo drive system is achieved through multi-sensor fusion, which can provide an effective decision-making basis for equipment maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular, to a method and system for monitoring the state of a servo drive system based on multi-sensor fusion. Background Art

[0002] In the fields of industrial automation and intelligent manufacturing, as a core power component, the operating state of the servo drive system directly affects the efficiency of the production line, product quality, and equipment safety. With the advent of the Industrial 4.0 era, higher requirements are put forward for the reliability and intelligent operation and maintenance level of the servo drive system. However, the existing state monitoring technologies for servo drive systems still face many challenges.

[0003] Related technologies often adopt static thresholds or simple statistical analysis methods, which are difficult to adapt to the dynamic changes of the equipment operating state. Such methods cannot adaptively extract the key features during the operation of the equipment, and it is even more difficult to predict the degradation trend of the equipment performance, resulting in a lag in operation and maintenance decisions and increasing the risk of unplanned downtime. At the same time, due to the lack of a multi-task collaborative evaluation mechanism, the existing technologies are difficult to comprehensively analyze the equipment health state, abnormal risk, and performance degradation trend at the same time, making the formulation of operation and maintenance strategies lack a global perspective and difficult to achieve precise and proactive maintenance operations. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for monitoring the state of a servo drive system based on multi-sensor fusion, the method comprising:

[0005] Obtain a multi-source sensor data set of a target servo drive device, the multi-source sensor data set including vibration waveform data, temperature gradient data, and current harmonic data;

[0006] Perform time synchronization and amplitude calibration processing on the multi-source sensor data set to generate a standardized monitoring data set;

[0007] Call a dynamic coding model to extract operation features from the standardized monitoring data set to generate an operation feature set of the target device;

[0008] Perform state evaluation on the operation feature set based on a multi-task state evaluation model to generate a real-time health state index, an abnormal risk level, and a performance degradation trend of the target device;

[0009] Generate an equipment maintenance optimization strategy according to the real-time health state index, the abnormal risk level, and the performance degradation trend, and feedback the equipment maintenance optimization strategy to an operation and maintenance control platform to trigger proactive maintenance operations.

[0010] In another aspect, an embodiment of the present invention further provides a state monitoring system for a servo drive system based on multi-sensor fusion, including a processor and a machine-readable storage medium. 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.

[0011] Based on the above aspects, the embodiments of the present application achieve global and in-depth perception and intelligent decision-making of the operating state of the target servo drive device, significantly improving the accuracy of device state monitoring and the intelligent level of operation and maintenance management. Specifically, by integrating multi-source heterogeneous sensor information such as vibration waveform data, temperature gradient data, and current harmonic data, and through time synchronization and amplitude calibration processing, the spatio-temporal differences and dimensional barriers between multi-source data are eliminated, ensuring the consistency and accuracy of data fusion. The introduction of the dynamic coding model enables the operation feature extraction process to adaptively capture the dynamic changes of the device state and generate a set of operation features with high discrimination, significantly improving the sensitivity and specificity of state evaluation. The multi-task state evaluation model is based on the above feature set to achieve synchronous evaluation of the real-time health status indicators, abnormal risk levels, and performance degradation trends of the device, not only revealing the current operating state of the device but also predicting potential failure risks and performance evolution trends. Finally, the device maintenance optimization strategy generated based on the evaluation results can accurately guide the operation and maintenance control platform to implement proactive maintenance operations, realizing the closed-loop optimization from state monitoring to operation and maintenance decision-making, effectively reducing the device failure rate, extending the service life of the device, and improving production efficiency and operation and maintenance economy at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic flowchart of the execution process of the state monitoring method for a servo drive system based on multi-sensor fusion provided by an embodiment of the present invention.

[0013] Figure 2 It is a schematic diagram of exemplary hardware and software components of a state monitoring system for a servo drive system based on multi-sensor fusion provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] The present invention will be specifically described below in conjunction with the accompanying drawings of the specification. Figure 1 It is a schematic flowchart of a state monitoring method for a servo drive system based on multi-sensor fusion provided by an embodiment of the present invention. The state monitoring method for a servo drive system based on multi-sensor fusion will be introduced in detail below.

[0015] Step S110: Obtain a multi-source sensor data set of a target servo drive device, where the multi-source sensor data set includes vibration waveform data, temperature gradient data, and current harmonic data.

[0016] In this embodiment, taking a servo drive device in a certain factory as an example, the servo drive device is responsible for driving multiple mechanical components to work together on the production line. To comprehensively monitor the operating state of the servo drive device, a variety of sensors are installed at key parts of the servo drive device. Specifically, a vibration sensor is installed on the motor housing to collect vibration waveform data to reflect the vibration situation during the operation of the servo drive device; temperature sensors are arranged near the motor winding and on the surfaces of some key transmission components to obtain temperature gradient data and understand the temperature changes of various parts of the servo drive device. The current sensor is installed on the power supply line of the servo drive device to collect current harmonic data in real time to analyze the electrical operating condition of the servo drive device.

[0017] Suppose at a certain moment, the vibration waveform data collected by the vibration sensor shows a series of peaks and valleys, and its data is recorded as [V1, V2, V3,..., Vn], where V1 represents the vibration amplitude at the first sampling moment, V2 represents the vibration amplitude at the second sampling moment, and so on, and n is the number of sampling points. The temperature gradient data collected by the temperature sensor, for example, the temperature sensor near the motor winding records the temperature values at certain time intervals, forming a temperature gradient data sequence [T1, T2, T3,..., Tm], T1 is the temperature value at the starting moment, Tm is the temperature value at the mth sampling moment, and m is the number of temperature sampling points. The current harmonic data collected by the current sensor is also recorded in sequence form, such as [I1, I2, I3,..., Ik], I1 is the current value at the first sampling moment, Ik is the current value at the kth sampling moment, and k is the number of current sampling points.

[0018] Step S120: Perform time synchronization and amplitude calibration processing on the multi-source sensor data set to generate a standardized monitoring data set.

[0019] In this embodiment, step S120 may include:

[0020] Step S121: Extract the first sampling timestamp sequence of the vibration waveform data, the second sampling timestamp sequence of the temperature gradient data, and the third sampling timestamp sequence of the current harmonic data.

[0021] In this embodiment, the acquisition system of vibration waveform data can record the corresponding timestamps for each sampling point, forming a first sampling timestamp sequence, such as [ts1, ts2, ts3, …, tsn], where ts1 represents the timestamp of the first vibration data sampling point, and tsn represents the timestamp of the nth vibration data sampling point. The acquisition system of temperature gradient data will also record the corresponding timestamps, forming a second sampling timestamp sequence, such as [tt1, tt2, tt3, …, ttm], where tt1 is the timestamp of the first temperature data sampling point, and ttm is the timestamp of the mth temperature data sampling point. The current harmonic data also has a corresponding third sampling timestamp sequence, such as [ti1, ti2, ti3, …, tik], where ti1 is the timestamp of the first current data sampling point, and tik is the timestamp of the kth current data sampling point.

[0022] Step S122: Determine the maximum common time interval among the first sampling timestamp sequence, the second sampling timestamp sequence, and the third sampling timestamp sequence, and perform time-axis alignment processing on the vibration waveform data, temperature gradient data, and current harmonic data based on the linear interpolation algorithm to generate a time-synchronized data set.

[0023] In this embodiment, after obtaining the first sampling timestamp sequence, the second sampling timestamp sequence, and the third sampling timestamp sequence, it is necessary to find the maximum common time interval among them. Suppose after analysis, it is found that the maximum common time interval is Δt. For the vibration waveform data, since its sampling frequency may be different from that of the temperature gradient data and the current harmonic data. For example, the sampling frequency of the vibration waveform data is fs1, the sampling frequency of the temperature gradient data is fs2, and the sampling frequency of the current harmonic data is fs3. If fs1 > fs2, then during the time-axis alignment process, for the temperature gradient data, new data points need to be inserted between the original sampling points through the linear interpolation algorithm to align it with the vibration waveform data in time. For example, between the time points tsi and tsi+1 of the vibration waveform data, according to the linear interpolation algorithm, for the temperature gradient data, the interpolated temperature value at the corresponding time point within this time period is calculated. Similarly, similar processing is also performed on the current harmonic data, and finally, the vibration waveform data, temperature gradient data, and current harmonic data are aligned on the time axis to form a time-synchronized data set.

[0024] Step S123: Call the preset set of sensor calibration coefficients to perform amplitude compensation processing on the vibration waveform data, temperature gradient data, and current harmonic data in the time-synchronized data set respectively, where the set of sensor calibration coefficients includes the frequency response compensation factor of the vibration sensor, the nonlinear error correction factor of the temperature sensor, and the phase shift correction parameter of the current sensor.

[0025] In this embodiment, the preset sensor calibration coefficient set can be obtained through a large number of experiments and calibrations. For vibration waveform data, assuming that the frequency response compensation factor of the vibration sensor is Cf1, within a certain frequency band, due to the frequency response characteristics of the sensor, there may be deviations in the collected vibration amplitudes. For example, under the vibration signal with a frequency of f1, the collected vibration amplitude is Va. According to the calibration coefficient Cf1, the vibration amplitude Vb after amplitude compensation is Vb = Va × Cf1. For temperature gradient data, the temperature sensor may have a non - linear error, and its non - linear error correction factor is Ct. Assuming that the collected temperature value is Ta, considering the non - linear error, the corrected temperature value Tb = Ta + Ct×(Ta - T0)², where T0 is the reference temperature value. For current harmonic data, the current sensor may have a phase shift, and its phase shift correction parameter is Ci. Assuming that the collected current value is Ia, the current value Ib after phase shift correction is Ib = Ia×exp(j×Ci). Here, the phase of the current is adjusted through the phase correction parameter to obtain a more accurate current amplitude. Through the above amplitude compensation processing operations, the amplitudes of the sensor data can be made more accurate and reliable.

[0026] Step S124: Normalize the vibration waveform data, temperature gradient data, and current harmonic data after amplitude compensation to generate a standardized monitoring data set.

[0027] In this embodiment, after amplitude compensation, the corresponding data can be normalized. Specifically, for vibration waveform data, assuming that the vibration data sequence after amplitude compensation is [Vb1, Vb2, Vb3, …, Vbn], its normalization formula is Vn_i=(Vb_i - min(Vb)) / (max(Vb)-min(Vb)), where Vn_i is the i - th vibration data after normalization, min(Vb) is the minimum value in the vibration data sequence after amplitude compensation, and max(Vb) is the maximum value. For temperature gradient data, the temperature data sequence after amplitude compensation is [Tb1, Tb2, Tb3, …, Tbm], and the normalization formula is Tn_j=(Tb_j - min(Tb)) / (max(Tb)-min(Tb)), where Tn_j is the j - th temperature data after normalization. For current harmonic data, the current data sequence after amplitude compensation is [Ib1, Ib2, Ib3, …, Ibk], and the normalization formula is In_k=(Ib_k - min(Ib)) / (max(Ib)-min(Ib)), where In_k is the k - th current data after normalization. Through the above normalization process, the amplitude ranges of the sensor data are unified to between [0, 1] to generate a standardized monitoring data set.

[0028] Step S130: Invoke the dynamic encoding model to extract the operation characteristics from the standardized monitoring data set, and generate an operation characteristics set of the target device.

[0029] In this embodiment, step S130 may include:

[0030] Step S131: Extract the time-domain characteristics from the vibration waveform data in the standardized monitoring data set to generate the vibration effective value, peak factor, and kurtosis coefficient.

[0031] For the vibration waveform data, taking the standardized vibration data sequence [Vn1, Vn2, Vn3, …, Vnn] as an example. The calculation method of the vibration effective value is to first square the amplitude of each sampling point, then find the average value of the above squared values, and finally take the square root. That is, the vibration effective value Veff = √((Vn1² + Vn2² + Vn3² + … + Vnn²) / n). The peak factor is the ratio of the vibration peak value to the effective value. First, find the maximum value Vmax in the vibration data sequence, then the peak factor CF = Vmax / Veff. The kurtosis coefficient is used to measure the impact characteristics of the vibration signal. Its calculation method is to first perform a fourth-power operation on the amplitude of each sampling point minus the average value, then find the average value of the above fourth-power values, and then divide by the fourth power of the effective value. Assuming the average value of the vibration data is Vavg, the kurtosis coefficient Kurtosis = (((Vn1 - Vavg) 4 + (Vn2 - Vavg) 4 + (Vn3 - Vavg) 4 + … + (Vnn - Vavg) 4 ) / n) / Veff 4 .

[0032] Step S132: Perform wavelet packet decomposition on the vibration waveform data to extract the preset frequency band energy ratio characteristics and frequency band entropy value characteristics.

[0033] Step S1321: Select the Daubechies wavelet basis function to perform 5-layer wavelet packet decomposition on the vibration waveform data to obtain 32 frequency band node coefficients.

[0034] In this embodiment, the Daubechies wavelet basis function is selected to decompose the vibration waveform data. Perform 5-layer wavelet packet decomposition on the standardized vibration waveform data. According to the principle of wavelet packet decomposition, 2 5 = 32 frequency band node coefficients will be obtained. For example, the frequency band node coefficients obtained after decomposition are [W1, W2, W3, …, W32], and each frequency band node coefficient represents the signal characteristics in a different frequency band.

[0035] Step S1322: Calculate the energy value of each frequency band node, and divide each frequency band energy value by the total energy to obtain the preset frequency band energy proportion feature.

[0036] For each frequency band node coefficient Wi, its energy value Ei = Wi². The total energy E = E1 + E2 + E3 + … + E32. The preset frequency band energy proportion feature Pi = Ei / E. For example, P1 = E1 / E, P2 = E2 / E, and so on. Thus, the energy proportion of each frequency band is obtained.

[0037] Step S1323: Calculate the permutation entropy for the coefficient sequence of each frequency band node to obtain the frequency band entropy value feature.

[0038] In this embodiment, for the coefficient sequence of each frequency band node, for example, the coefficient sequence [Wi1, Wi2, Wi3, …, Win] of the i-th frequency band node, calculate its permutation entropy. First, determine the embedding dimension m and the delay time τ. Assume m = 3 and τ = 1. Reconstruct the coefficient sequence into an m-dimensional vector. For example, for the frequency band node with i = 1, the reconstructed vector is [(W11, W12, W13), (W12, W13, W14), …]. Then sort the above vectors, count the number of occurrences of different permutation orders, and calculate the permutation entropy of this frequency band node according to the permutation entropy calculation formula, denoted as Si. Perform such calculations for all 32 frequency band nodes to obtain the frequency band entropy value feature.

[0039] Step S1324: Screen the characteristic frequency bands whose energy proportion exceeds the set threshold, and add the corresponding frequency band entropy value feature and energy proportion feature to the running feature set.

[0040] In this embodiment, assume that the set energy proportion threshold is Th, and screen the energy proportion features of the 32 calculated frequency bands. For example, it is found that the energy proportions P5, P8, P12 of the 5th, 8th, 12th, etc. frequency bands exceed the threshold Th. Then add the corresponding frequency band entropy value features S5, S8, S12 and energy proportion features P5, P8, P12 of the above frequency bands to the running feature set.

[0041] Step S133: Perform a sliding window mean process on the temperature gradient data to generate a temperature change rate feature and a heat accumulation trend feature.

[0042] In this embodiment, for the standardized temperature gradient data sequence [Tn1, Tn2, Tn3, …, Tnm], perform a sliding window mean process. Assume that the sliding window size is w, for example, w = 5. For the temperature change rate feature, take the i-th window as an example. The data within the window is [Tn i , Tn i+1 , Tn i+2 , Tn i+3 , Tn i+4, the temperature change rate feature Rt_i = (Tn i+4 - Tn i ) / 4. By moving the window, a series of temperature change rate feature values [Rt1, Rt2, Rt3, …] are obtained. For the heat accumulation trend feature, calculate the sum of the temperature values within each window. For example, the heat accumulation value Ht_i of the i-th window is Ht_i = Tn i + Tn i+1 + Tn i+2 + Tn i+3 + Tn i+4 , and by moving the window, a heat accumulation trend feature sequence [Ht1, Ht2, Ht3, …] is obtained.

[0043] Step S134: Perform fast Fourier transform processing on the current harmonic data to extract the fundamental wave amplitude, harmonic distortion rate, and amplitude ratio of characteristic sub-harmonic components.

[0044] In this embodiment, perform fast Fourier transform processing on the standardized current harmonic data sequence [In1, In2, In3, …, Ink]. After fast Fourier transform, frequency domain data is obtained. In the frequency domain, find the amplitude corresponding to the fundamental wave frequency, denoted as Ifundamental. The calculation method of the harmonic distortion rate THD is to first calculate the sum of the squares of the amplitudes of all harmonics except the fundamental wave, then divide it by the square of the fundamental wave amplitude, and then take the square root and multiply by 100%. That is, THD = √((Ih1² + Ih2² + Ih3² + …) / Ifundamental²) × 100%, where Ih1, Ih2, Ih3, etc. are the amplitudes of each harmonic. For the amplitude ratio of characteristic sub-harmonic components, assuming that the 3rd and 5th harmonics are concerned, the amplitude ratio of characteristic sub-harmonic components R = Ih3 / Ih5.

[0045] Step S135: Input the vibration effective value, peak factor, kurtosis coefficient, preset frequency band energy ratio feature, frequency band entropy value feature, temperature change rate feature, heat accumulation trend feature, fundamental wave amplitude, harmonic distortion rate, and amplitude ratio of characteristic sub-harmonic components into the dynamic encoding model, and perform dynamic correlation encoding through a multi-layer cross-attention mechanism to generate an operating feature set.

[0046] In this embodiment, all the features extracted previously, namely the effective vibration value Veff, crest factor CF, kurtosis coefficient Kurtosis, preset band energy ratio features [P5, P8, P12], band entropy value features [S5, S8, S12], temperature change rate features [Rt1, Rt2, Rt3], heat accumulation trend features [Ht1, Ht2, Ht3], fundamental wave amplitude Ifundamental, harmonic distortion rate THD, and characteristic sub-harmonic component amplitude ratio R are input into the dynamic coding model. The dynamic coding model performs dynamic correlation coding on the above different types of features through a multi-layer cross-attention mechanism. For example, in the first layer of the attention mechanism, the model analyzes the correlation between vibration features and temperature features and performs weighted fusion according to their potential relationships during the operation of the device. In the second layer of the attention mechanism, the connection between electrical features (such as current harmonic features) and the features fused previously is further considered, and weighted fusion and coding are performed again. Finally, a set of operating features of the target device is generated, and this set of operating features contains multi-dimensional feature information after comprehensive analysis and coding.

[0047] Step S140: Based on the multi-task state evaluation model, perform state evaluation on the set of operating features to generate the real-time health status index, abnormal risk level, and performance degradation trend of the target device.

[0048] In this embodiment, step S140 may include:

[0049] Step S141: Input the set of operating features into the shared feature extraction layer to generate a shared hidden feature vector.

[0050] Specifically, the set of operating features contains feature information in multiple dimensions and is input into the shared feature extraction layer of the multi-task state evaluation model. The shared feature extraction layer is a convolutional neural network structure that can perform feature extraction and compression on the input set of operating features. For example, the set of operating features can be understood as a multi-dimensional feature matrix, and the shared feature extraction layer performs sliding convolution operations on this matrix through convolutional kernels to extract the key feature information. After operations such as convolution and pooling, the high-dimensional set of operating features is compressed into a low-dimensional shared hidden feature vector, and this shared hidden feature vector contains the key information in the set of operating features.

[0051] Step S142: Input the shared hidden feature vector into the real-time health status prediction branch, abnormal risk classification branch, and performance degradation regression branch respectively.

[0052] In this embodiment, after the shared implicit feature vector is generated, it is respectively input into three different task branches. Specifically, the real-time health status prediction branch is responsible for predicting the current health status of the device; the abnormal risk classification branch is used to judge the level of abnormal risk faced by the device currently; and the performance degradation regression branch quantifies and evaluates the performance degradation trend of the device.

[0053] Step S143: In the real-time health status prediction branch, based on the gated recurrent unit network, perform temporal dependence modeling on the shared implicit feature vector, and output the real-time health status index.

[0054] In this embodiment, in the real-time health status prediction branch, the gated recurrent unit network (GRU) is used to process the shared implicit feature vector. The GRU network can capture the temporal dependence relationship in the time series data. Assuming that the shared implicit feature vector is H, the GRU network will update the hidden state at the current moment according to the input at the current moment and the hidden state at the previous moment. For example, at time t, the input of the GRU network is H_t, and the hidden state at the previous moment is h_t-1. Through the calculation formula of the GRU network, the hidden state h_t at the current moment is updated. After being processed through multiple time steps, a real-time health status index is finally output. This real-time health status index can be a numerical value, for example, between 0 and 100, representing the current health degree of the device. The higher the value, the better the health status.

[0055] Step S144: In the abnormal risk classification branch, based on the multi-head self-attention mechanism, focus on the key features of the shared implicit feature vector, and output the probability distribution of the abnormal risk level.

[0056] In this embodiment, the multi-head self-attention mechanism is used in the abnormal risk classification branch. The multi-head self-attention mechanism can simultaneously focus on different parts of the shared implicit feature vector to focus on the key features. For the shared implicit feature vector H, the multi-head self-attention mechanism will project it into multiple subspaces, for example, divided into 8 heads. Each head will calculate the attention weights between the feature vectors and focus on the key features through the above weights. For example, one head may pay more attention to the features related to vibration abnormalities, and another head may pay more attention to the features related to temperature abnormalities. After being processed by the multi-head self-attention mechanism, a probability distribution vector is output. Each element of this vector represents the probability that the device is in different abnormal risk levels. For example, the probability distribution vector is [P1, P2, P3, P4], where P1 represents the probability that the device is in the low risk level, P2 represents the probability of being in the medium risk level, P3 represents the probability of being in the high risk level, and P4 represents the probability of being in the extremely high risk level.

[0057] Step S145: In the performance degradation regression branch, perform a non-linear mapping on the shared hidden feature vector based on a radial basis function network, and output a quantization value of the performance degradation trend.

[0058] In this embodiment, the performance degradation regression branch uses a radial basis function network (RBF network) to process the shared hidden feature vector. The RBF network can perform a non-linear mapping on the input feature vector. For the shared hidden feature vector H, the RBF network uses a radial basis function as the activation function, such as the commonly used Gaussian radial basis function. Assuming that the input shared hidden feature vector H is a multi-dimensional vector [x1, x2, x3,..., xn], the RBF network will calculate the distance between this vector and each center vector ci, for example, the distance di = ||H - ci|| (where ||.|| represents a certain norm, such as the Euclidean norm). Then, through the Gaussian radial basis function, such as exp(-di² / (2σ²)), where σ is the width parameter, the distance is mapped to an activation value. The RBF network performs a non-linear transformation on the input through the above activation value, and finally outputs a quantization value of the performance degradation trend through a linear combination. This quantization value can represent the degree of degradation of the device performance over time. For example, if the quantization value is positive, it means the performance is gradually degrading, and the larger the quantization value, the faster the degradation speed; if the quantization value is negative or close to 0, it means the performance is relatively stable or has an improving trend.

[0059] Step S146: Dynamically adjust the weight allocation ratio of the shared feature extraction layer according to the output results of the real-time health status prediction branch, the abnormal risk classification branch, and the performance degradation regression branch, so as to achieve multi-task collaborative optimization.

[0060] In this embodiment, Step S146 may include:

[0061] Step S1461: Calculate the prediction error rate of the real-time health status prediction branch, the cross-entropy loss value of the abnormal risk classification branch, and the mean square error of the performance degradation regression branch.

[0062] For the real-time health status prediction branch, assuming that the real-time health status metric output is y_pred, and the actual health status label is y_true, the prediction error rate is calculated by first calculating the absolute error between the predicted value and the true value, and then finding the ratio of the average of the above errors to the average of the true values. That is, the prediction error rate E1 = mean(|y_pred - y_true|) / mean(y_true). For the abnormal risk classification branch, what it outputs is a probability distribution vector P_pred, and the actual abnormal risk level label is a one-hot encoded vector P_true (for example, if the device is in the medium risk level, the element corresponding to the medium risk level in P_true is 1, and other elements are 0), and the cross-entropy loss value E2 = -Σ(P_true[i]*log(P_pred[i])), where i iterates over all possible risk level categories. For the performance degradation regression branch, assuming that the quantified value of the performance degradation trend output is z_pred, and the actual performance degradation trend label is z_true, the mean squared error E3 = mean((z_pred - z_true)²).

[0063] Step S1462: Normalize the prediction error rate, cross-entropy loss value, and mean squared error to obtain the weight adjustment factors for each branch.

[0064] To make different error metrics comparable, it is necessary to normalize the calculated prediction error rate E1, cross-entropy loss value E2, and mean squared error E3. Assume that the normalized prediction error rate is E1_norm = (E1 - min(E1)) / (max(E1) - min(E1)), the cross-entropy loss value after normalization is E2_norm = (E2 - min(E2)) / (max(E2) - min(E2)), and the mean squared error after normalization is E3_norm = (E3 - min(E3)) / (max(E3) - min(E3)). The above normalized error values are the weight adjustment factors for each branch, denoted as w1 = E1_norm, w2 = E2_norm, and w3 = E3_norm respectively.

[0065] Step S1463: Dynamically update the channel attention weights of each convolution kernel in the shared feature extraction layer according to the weight adjustment factors.

[0066] In this embodiment, there are multiple convolutional kernels in the shared feature extraction layer, and each convolutional kernel is responsible for extracting different feature information. According to the weight adjustment factor, the channel attention weights of the convolutional kernels are dynamically updated. For example, for the j-th channel of a certain convolutional kernel, its original attention weight is a_j, and the updated weight a_j' = a_j * (w1 + w2 + w3) / 3 (this is just a simple weighted update method, and in practice, more complex adjustments may be made according to specific algorithms). In this way, the shared feature extraction layer can dynamically adjust the degree of attention to different features according to the performance of each branch task, so as to achieve multi-task collaborative optimization.

[0067] Step S1464: When the weight adjustment factor of the abnormal risk classification branch exceeds the preset alarm threshold, freeze the parameter update of the performance degradation regression branch and preferentially optimize the abnormal risk classification task.

[0068] In this embodiment, it is assumed that the preset alarm threshold is Threshold. When the weight adjustment factor w2 of the abnormal risk classification branch exceeds Threshold, it indicates that the error of the abnormal risk classification task is relatively large and needs to be optimized preferentially. At this time, freeze the parameter update of the performance degradation regression branch, that is, no longer adjust its network parameters according to the error of the performance degradation regression branch. This can concentrate computing resources and optimization directions, preferentially improve the accuracy of abnormal risk classification, and ensure that abnormal situations of the device can be detected and processed in time.

[0069] Step S150: Generate a device maintenance optimization strategy according to the real-time health status indicator, abnormal risk level, and performance degradation trend, and feedback the device maintenance optimization strategy to the operation and maintenance control platform to trigger proactive maintenance operations.

[0070] In this embodiment, step S150 may include;

[0071] Step S151: Establish a threshold interval for the health status indicator, a threshold interval for the abnormal risk level, and a slope threshold for the performance degradation trend.

[0072] In this embodiment, through a large number of experiments and statistical analysis of equipment operation data, a corresponding threshold range is established. The threshold range of the health status index is set as [Health_min, Health_max]. For example, Health_min = 60 and Health_max = 80, indicating that when the real-time health status index is between 60 and 80, the equipment is in a relatively normal health state. The threshold ranges of the abnormal risk levels are divided into thresholds of different levels. For example, the low-risk level threshold is Risk_low = 0.3, the medium-risk level threshold is Risk_mid = 0.6, and the high-risk level threshold is Risk_high = 0.8. The threshold of the performance degradation trend slope is set as Slope_threshold = 0.05, indicating that when the performance degradation trend slope exceeds 0.05, the performance degradation of the equipment needs to be concerned about.

[0073] Step S152: When the real-time health status index is lower than the lower limit of the threshold range of the health status index and the abnormal risk level exceeds the first risk threshold, generate an immediate shutdown and maintenance instruction.

[0074] In this embodiment, assume that the real-time health status index is Health_index and the abnormal risk level is Risk_level. When Health_index < Health_min and Risk_level > Risk_low, it indicates that the health condition of the equipment is poor and it faces a relatively high abnormal risk. For example, Health_index = 50 and Risk_level = 0.4. At this time, the system will generate an immediate shutdown and maintenance instruction to avoid further damage to the equipment and ensure production safety.

[0075] Step S153: When the performance degradation trend slope exceeds the slope threshold and the abnormal risk level is within the second risk threshold range, generate a preventive maintenance plan and spare part replacement suggestions.

[0076] In this embodiment, assume that the performance degradation trend slope is Slope. When Slope > Slope_threshold and Risk_low <= Risk_level <= Risk_mid, for example, Slope = 0.06 and Risk_level = 0.5, at this time, a preventive maintenance plan and spare part replacement suggestions need to be generated.

[0077] Step S1531: Calculate the predicted value of the remaining service life according to the performance degradation trend slope.

[0078] In this embodiment, based on the slope of the performance degradation trend, the historical operation data of the device, and the performance model, the predicted remaining useful life of the device can be calculated. For example, by establishing a linear model, assuming the initial performance of the device is P0, the current performance is P1, the slope of the performance degradation trend is Slope, and the rated useful life of the device is T0. Then the predicted remaining useful life Remaining_life = (P0 - P1) / Slope. Assume P0 = 100 (indicating the initial full-performance state of the device), P1 = 80, and Slope = 0.06, then Remaining_life = (100 - 80) / 0.06 ≈ 333 (the unit can be hours, days, etc., assumed to be hours here).

[0079] Step S1532: Query the device maintenance knowledge graph and match the set of historical maintenance cases corresponding to the predicted remaining useful life value.

[0080] In this embodiment, a large number of historical device maintenance cases are stored in the device maintenance knowledge graph. According to the calculated predicted remaining useful life value, query and match in the device maintenance knowledge graph. For example, a set of device maintenance cases with a remaining useful life between 300 - 350 hours in history is found. The above device maintenance cases contain maintenance measures and experiences of different devices under similar performance degradation conditions.

[0081] Step S1533: Extract the optimal maintenance time window, spare part model replacement records, and labor cost data from the historical maintenance cases.

[0082] In this embodiment, the optimal maintenance time window, spare part model replacement records, and labor cost data of each case can be extracted from the matched set of historical maintenance cases. For example, in a certain historical case, the optimal maintenance time window is 50 - 80 hours before the remaining useful life, the spare part model replacement record is to replace the brush of the motor and a certain bearing, and the labor cost is 5000 yuan.

[0083] Step S1534: Based on the fuzzy decision-making algorithm, perform multi-objective optimization on the maintenance time window, spare part inventory status, and production plan to generate the time nodes of the preventive maintenance plan, the list of required spare parts, and the human resource allocation plan.

[0084] In this embodiment, step S1534 may include:

[0085] Step S15341: Convert the maintenance time window into a time overlap conflict matrix, map the spare part inventory status to a spare part availability vector, and parse the production plan into a maintenance operation time constraint sequence.

[0086] In this embodiment, for the maintenance time window, assuming that multiple maintenance time windows obtained from historical cases are [Time_window1, Time_window2, …], they are converted into a time overlap conflict matrix. For example, if Time_window1 = [t1, t2] and Time_window2 = [t3, t4], by analyzing whether there is an overlap between the above time windows, a matrix is constructed, and the elements in the matrix represent the overlap situation between different time windows. For the spare part inventory status, assuming there are multiple spare parts in the inventory, namely Spare_part1, Spare_part2, …, they are mapped to a spare part availability vector, such as [Available1, Available2, …], where Available1 represents the available quantity of Spare_part1. For the production plan, the maintenance operation time constraint sequence is parsed. For example, the production plan requires that the equipment cannot be shut down for maintenance during a certain time period, and the above constraints are sorted into a sequence.

[0087] Step S15342: Based on the time overlap conflict matrix, spare part availability vector, and maintenance operation time constraint sequence, construct a multi-objective optimization space, and use the Pareto front search algorithm to generate a set of candidate maintenance plans.

[0088] In this embodiment, a multi-objective optimization space can be constructed using the time overlap conflict matrix, spare part availability vector, and maintenance operation time constraint sequence. The objectives in this multi-objective optimization space include minimizing the conflict between the maintenance time window and the production plan, maximizing the utilization rate of spare parts, and minimizing the labor cost, etc. Using the Pareto front search algorithm, a set of candidate maintenance plans is searched in this multi-objective optimization space. The above candidate maintenance plans achieve a balance between different objectives, and there is no situation where one plan is superior to other plans in all objectives.

[0089] Step S15343: Perform time conflict resolution processing on the set of candidate maintenance plans, eliminate the candidate plans that conflict with the maintenance operation time constraint sequence, and generate a set of conflict-free candidate maintenance plans.

[0090] In this embodiment, for each plan in the set of candidate maintenance plans, check whether its maintenance time conflicts with the maintenance operation time constraint sequence. For example, if the maintenance time of a certain candidate plan overlaps with the non-stop operation time period specified in the production plan, then this plan is eliminated. After such processing, a set of conflict-free candidate maintenance plans is generated.

[0091] Step S15344: According to the spare part availability vector, sort the set of conflict-free candidate maintenance plans by the spare part matching degree, and select the top k candidate plans with the highest spare part matching degree.

[0092] In this embodiment, the matching degree between the spare parts required in each conflict-free candidate maintenance plan and the in-stock spare parts can be calculated according to the spare part availability vector. For example, if a certain candidate plan requires 5 pieces of Spare_part1 and the available quantity of Spare_part1 in stock is 8, then the matching degree of this plan for Spare_part1 is relatively high. By sorting the conflict-free candidate maintenance plans according to the spare part matching degree, the top k candidate plans with the highest spare part matching degree are selected.

[0093] Step S15345: Input the top k candidate plans into the fuzzy decision-making device, and generate the time nodes, required spare part list, and human resource allocation plan of the preventive maintenance plan according to the preset maintenance priority rules.

[0094] In this embodiment, the selected top k candidate plans can be input into the fuzzy decision-making device. The fuzzy decision-making device comprehensively evaluates and makes decisions on the above plans according to the preset maintenance priority rules, such as giving priority to the continuity of the production plan and minimizing costs. Finally, the time nodes of the preventive maintenance plan are generated, for example, it is determined to perform maintenance 60 hours before the remaining service life; the required spare part list, such as listing spare parts that need to be replaced, such as motor brushes, bearings, etc.; and the human resource allocation plan, such as arranging several technicians to participate in the maintenance work, etc.

[0095] Step S154: When the real-time health status indicator is located in the middle of the health status indicator threshold range and the abnormal risk level is lower than the third risk threshold, generate an operation parameter optimization and adjustment plan.

[0096] In this embodiment, when Health_min < Health_index < Health_max and Risk_level < Risk_mid, for example, Health_index = 70 and Risk_level = 0.2, it means that the current health status of the equipment is relatively good and the abnormal risk is low. At this time, an operation parameter optimization and adjustment plan is generated. By adjusting some operation parameters of the equipment, such as the speed and voltage of the motor, the operation efficiency and stability of the equipment are further improved. For example, according to the operation characteristics and historical data of the equipment, it is found that when the motor speed is adjusted from the current 1500 revolutions per minute to 1450 revolutions per minute, the energy consumption of the equipment can be reduced by 5% without affecting the production quality. Then the operation parameter optimization and adjustment plan may include measures such as adjusting the motor speed to 1450 revolutions per minute and corresponding voltage adjustment.

[0097] Step S155: Integrate the immediate shutdown for maintenance instruction, preventive maintenance plan, spare part replacement suggestion, and operation parameter optimization and adjustment plan into an equipment maintenance optimization strategy, and feedback the equipment maintenance optimization strategy to the operation and maintenance control platform to trigger proactive maintenance operations.

[0098] In this embodiment, the generation of an immediate shutdown and maintenance instruction, a preventive maintenance plan, a spare part replacement suggestion, and an operation parameter optimization and adjustment plan can be integrated to form a complete equipment maintenance and optimization strategy. Then, this strategy is fed back to the operation and maintenance control platform. After receiving the strategy, the operation and maintenance control platform will trigger corresponding proactive maintenance operations according to the content of the strategy. For example, if it is an immediate shutdown and maintenance instruction, the operation and maintenance control platform will send an instruction to the equipment control system to stop the equipment immediately and notify the maintenance personnel to carry out maintenance; if it is a preventive maintenance plan and a spare part replacement suggestion, the operation and maintenance control platform will arrange for the maintenance personnel to carry out maintenance work at a specified time and prepare the required spare parts; if it is an operation parameter optimization and adjustment plan, the operation and maintenance control platform will send an instruction to adjust the parameters to the controller of the equipment to achieve the adjustment of the equipment operation parameters.

[0099] For example, in a further implementation, the method may further include a training step of a multi-task state evaluation model:

[0100] Step S210: Obtain a multi-task training data set, where the multi-task training data set includes a historical operation feature set and its corresponding health status label, abnormal risk level label, and performance degradation trend label.

[0101] For example, step S210 may include:

[0102] Step S2101: Collect multi-source sensor monitoring data of the servo drive device during its complete life cycle, and perform time synchronization and amplitude calibration processing to generate a standardized historical monitoring data set.

[0103] Specifically, during the entire life cycle of the servo drive device, continuously collect its multi-source sensor monitoring data. For example, starting from the equipment installation and commissioning stage, vibration sensors, temperature sensors, and current sensors continuously collect data. At different stages of equipment operation, such as normal operation, overload operation, and fault occurrence, corresponding data records are available. The collected data also needs to be subjected to time synchronization and amplitude calibration processing, and the processing process is the same as the processing method when obtaining real-time data before. For example, for vibration waveform data, temperature gradient data, and current harmonic data, respectively extract the sampling timestamp sequence, determine the maximum common time interval, align the time axis, then call the preset sensor calibration coefficient set for amplitude compensation, and finally perform normalization processing to generate a standardized historical monitoring data set.

[0104] Step S2102: Extract operation features from the standardized historical monitoring data set to generate a historical operation feature set.

[0105] In this embodiment, the operating characteristics of the standardized historical monitoring data set can be extracted, and the process is similar to that of the operating characteristics extraction of real-time data. For example, for vibration waveform data, time-domain characteristics are extracted, and the effective value of vibration, peak factor, and kurtosis coefficient are calculated; wavelet packet decomposition is performed to extract the preset band energy ratio characteristics and band entropy value characteristics. For temperature gradient data, sliding window mean processing is performed to generate temperature change rate characteristics and heat accumulation trend characteristics. For current harmonic data, fast Fourier transform processing is performed to extract the fundamental wave amplitude, harmonic distortion rate, and amplitude ratio of characteristic sub-harmonic components. The above characteristics are combined together to form a historical operating characteristics set.

[0106] Step S2103: Based on the equipment maintenance records, label the health status label for each time point, and the health status label is quantified according to the ratio of the actual number of repairs to the standard maintenance cycle.

[0107] In this embodiment, the equipment maintenance records detail the maintenance conditions of the equipment at different time points. According to the above records, calculate the ratio of the actual number of repairs to the standard maintenance cycle for each time point to quantify the health status label. For example, the standard maintenance cycle stipulates that the equipment undergoes a comprehensive maintenance every 1000 hours. At a certain time point, the equipment has been running for 5000 hours and has actually been repaired 3 times. Then the quantified value of the health status label at this time point is 3 / (5000 / 1000)=0.6. The value of the health status label is between 0 and 1, and the smaller the value, the better the health status of the equipment.

[0108] Step S2104: Based on the fault event records, label the abnormal risk level label, and the abnormal risk level label is graded according to the standard deviation of the characteristic fluctuation amplitude within N hours before the fault occurs.

[0109] In this embodiment, for the fault event records, analyze the fluctuation conditions of various characteristics of the equipment within N hours before the fault occurs. For example, select N = 24 hours and analyze the fluctuation amplitudes of characteristics such as vibration amplitude, temperature change, and current harmonics within these 24 hours. Calculate the standard deviation of the above characteristic fluctuation amplitudes, and perform abnormal risk level grading according to the size of the standard deviation. For example, when the standard deviation is less than a certain threshold, it is labeled as a low risk level; when the standard deviation is within a certain range, it is labeled as a medium risk level; when the standard deviation is greater than a certain higher threshold, it is labeled as a high risk level.

[0110] Step S2105: Based on the performance test data, label the performance degradation trend label, and the performance degradation trend label is obtained by linear fitting according to the percentage of the output torque decay rate to the rated torque.

[0111] In this embodiment, the performance data of the device, such as the output torque, can be obtained by periodically performing performance tests on the device. Analyze the attenuation of the output torque over time, and calculate the percentage of the output torque attenuation rate to the rated torque. For example, if the rated torque of the device is 1000 N·m, and over a period of time, the output torque drops from the initial 1000 N·m to 950 N·m after 500 hours, then the output torque attenuation rate is (1000 - 950) / 500 = 0.1 N·m / hour. The percentage to the rated torque is 0.1 / 1000×100% = 0.01%. Based on a series of such data points, linear fitting is performed to obtain the performance degradation trend label. For example, after multiple performance tests and data fittings, a linear equation is obtained, which can represent the relationship between the output torque attenuation rate and time, and the relevant parameters or slope and other information of this linear equation are used as the performance degradation trend label. Through the above steps, a multi-task training dataset containing the historical operation feature set and its corresponding health status label, abnormal risk level label, and performance degradation trend label is obtained.

[0112] Step S220: Initialize the convolutional neural network parameters of the shared feature extraction layer, the gated recurrent unit network parameters of the real-time health status prediction branch, the multi-head self-attention mechanism parameters of the abnormal risk classification branch, and the radial basis function network parameters of the performance degradation regression branch.

[0113] In this embodiment, for the convolutional neural network of the shared feature extraction layer, its parameters include the weights and biases of the convolutional kernels. For example, assuming the convolutional kernel size is 3×3 and there are 16 convolutional kernels, then each convolutional kernel has 3×3 weight values and one bias value. The above parameters are assigned random values during initialization, but usually follow a certain distribution, such as a normal distribution or a uniform distribution. For example, the weight values are randomly taken between -0.1 and 0.1, and the bias value is initialized to 0.

[0114] The gated recurrent unit network parameters of the real-time health status prediction branch include the weights and biases of the update gate, reset gate, and output gate. For example, the weight matrix Wz of the update gate has a size of [input dimension, hidden dimension], the weight matrix Wr of the reset gate also has a size of [input dimension, hidden dimension], and the weight matrix Wh of the output gate has a size of [input dimension, hidden dimension]. Their weight values are also randomly taken during initialization, and the bias values are also initialized to appropriate small values.

[0115] The multi-head self-attention mechanism parameters of the abnormal risk classification branch include the projection matrix and attention weights. For example, assuming there are 8 heads, the projection matrix of each head projects the input feature vector into different subspaces, and the weights of the projection matrix are randomly assigned during initialization. The attention weights are also given small random values during initialization so as to be adjusted according to the data during the training process.

[0116] The radial basis function network parameters of the performance degradation regression branch include the center vector and the width parameter. For each radial basis function, the dimension of the center vector is the same as that of the input feature vector, and its value can be reasonably set according to the range of the input data during initialization, for example, taking values near the mean of the input data. The width parameter determines the width of the radial basis function and also sets a suitable value during initialization, such as 0.5. Through such initialization, an initial parameter setting is provided for the training of the multi-task state evaluation model.

[0117] Step S230: Input the historical operation feature set into the initialized shared feature extraction layer to generate a historical shared hidden feature vector.

[0118] In this embodiment, the obtained historical operation feature set can be used as the input and fed into the initialized shared feature extraction layer. The convolutional neural network of the shared feature extraction layer starts to process the historical operation feature set. For example, the historical operation feature set can be regarded as a multi-dimensional array containing various feature information such as vibration, temperature, and current. The convolutional kernels of the convolutional neural network perform sliding convolutional operations on this multi-dimensional array. Each convolution calculates the dot product of the convolutional kernel and the local features and adds a bias value. For example, for a 3×3 convolutional kernel, it performs a convolutional calculation with a 3×3 local feature region at a certain position to obtain a new feature value. After multiple convolutional operations, the feature information is gradually extracted and compressed. Then, a pooling operation may be performed, such as max pooling or average pooling, to further reduce the feature dimension and retain key features. For example, in max pooling, the maximum value in a local region is selected as the output after pooling. After a series of convolutional and pooling operations, a historical shared hidden feature vector is finally generated, which contains the key information in the historical operation feature set.

[0119] Step S240: Input the historical shared hidden feature vector into each branch network respectively, and calculate the mean square error between the output of the real-time health state prediction branch and the health state label, the cross-entropy loss between the output of the abnormal risk classification branch and the abnormal risk level label, and the cosine similarity loss between the output of the performance degradation regression branch and the performance degradation trend label, to obtain the loss values of each branch.

[0120] In this embodiment, the generated historical shared hidden feature vector can be input into the real-time health state prediction branch, the abnormal risk classification branch, and the performance degradation regression branch respectively.

[0121] In the real-time health status prediction branch, the gated recurrent unit network performs temporal dependence modeling based on the historical shared hidden feature vector and outputs a predicted real-time health status metric. Suppose the predicted value is y_pred and the corresponding health status label is y_true. The mean squared error is calculated by first computing the squared error between the predicted value and the true value for each sample, i.e., (y_pred[i] - y_true[i])², where i represents the sample index. Then, the sum of the squared errors for all samples is calculated and divided by the number of samples to obtain the mean squared error MSE = Σ(y_pred[i] - y_true[i])² / n, where n is the number of samples.

[0122] In the abnormal risk classification branch, after the multi-head self-attention mechanism focuses on the key features of the historical shared hidden feature vector, it outputs a probability distribution vector representing the probabilities of the device being in different abnormal risk levels. Suppose the output probability distribution vector is P_pred and the actual abnormal risk level label is a one-hot encoded vector P_true. The cross-entropy loss is calculated by computing -P_true[i] * log(P_pred[i]) for each category and then summing over all categories, i.e., the cross-entropy loss CE = -Σ(P_true[i] * log(P_pred[i])).

[0123] In the performance degradation regression branch, after the radial basis function network performs a non-linear mapping on the historical shared hidden feature vector, it outputs a quantified value of the performance degradation trend. Suppose the output value is z_pred and the performance degradation trend label is z_true. The cosine similarity loss is calculated by first computing the dot product of the predicted value vector and the true value vector, then dividing it by the product of their magnitudes, and finally subtracting this cosine similarity value from 1. That is, the cosine similarity Cosine_similarity = (z_pred · z_true) / (||z_pred|| ||z_true||), and the cosine similarity loss CL = 1 - Cosine_similarity. Through the above calculations, the loss values of each branch are obtained, and these loss values reflect the gap between the current prediction performance of each branch network and the true labels.

[0124] Step S250: Dynamically adjust the parameter update amplitude of the shared feature extraction layer according to the backpropagation gradients of the loss values of each branch, and preferentially optimize the network parameters corresponding to the branch task with the highest loss value.

[0125] For example, step S250 may include:

[0126] Step S2501: In each training iteration, input the loss values of each branch into the dynamic weight allocator to generate the gradient weighting coefficients of each branch.

[0127] In this embodiment, during each training iteration, the mean squared error of the real-time health status prediction branch, the cross-entropy loss of the anomaly risk classification branch, and the cosine similarity loss of the performance degradation regression branch are input into the dynamic weight allocator. The dynamic weight allocator calculates the gradient weighting coefficients of each branch according to the magnitudes of the above loss values. For example, a simple method can be adopted to calculate the proportion of the loss value of each branch in the total loss value as the gradient weighting coefficient. Suppose the mean squared error of the real-time health status prediction branch is E1, the cross-entropy loss of the anomaly risk classification branch is E2, and the cosine similarity loss of the performance degradation regression branch is E3, and the total loss value E = E1 + E2 + E3. Then the gradient weighting coefficient w1 of the real-time health status prediction branch is w1 = E1 / E, the gradient weighting coefficient w2 of the anomaly risk classification branch is w2 = E2 / E, and the gradient weighting coefficient w3 of the performance degradation regression branch is w3 = E3 / E. The above gradient weighting coefficients reflect the relative importance of the losses of each branch in the total loss.

[0128] Step S2502: Adjust the gradient magnitude of the parameters of the shared feature extraction layer during the backpropagation process according to the gradient weighting coefficients, so that the branch task with a higher loss value obtains a larger parameter update weight.

[0129] In this embodiment, during the backpropagation process, the parameter update of the shared feature extraction layer depends on the gradient information of each branch. According to the calculated gradient weighting coefficients, the gradient magnitude of the parameters of the shared feature extraction layer is adjusted. For example, for the weight parameter of a certain convolution kernel in the shared feature extraction layer, its original gradient is g. During adjustment, its gradient is multiplied by the corresponding gradient weighting coefficient. If the loss value of the anomaly risk classification branch is relatively high and its gradient weighting coefficient w2 is relatively large, then when the gradient related to the anomaly risk classification branch is backpropagated to the shared feature extraction layer, the corresponding gradient magnitude will be amplified. In this way, the branch task with a higher loss value can obtain a larger parameter update weight in the shared feature extraction layer, so as to preferentially optimize the network parameters corresponding to this branch task and improve its prediction performance.

[0130] Step S2503: When the cross-entropy loss of the anomaly risk classification branch has not decreased for N consecutive iterations, freeze the parameter update of the real-time health status prediction branch to strengthen the feature extraction ability of the anomaly risk classification task.

[0131] In this embodiment, a threshold N is set, for example, N=5. When the cross entropy loss of the abnormal risk classification branch does not decrease in five consecutive iterations, it means that the current training process may not effectively optimize the abnormal risk classification task. At this time, in order to concentrate resources to optimize the abnormal risk classification task, the parameter update of the real-time health status prediction branch is frozen. That is, the gated recurrent unit network parameters of the real-time health status prediction branch are no longer adjusted according to the current error. In this way, the parameter update of the real-time health status prediction branch can be prevented from interfering with the optimization process of the abnormal risk classification task, so that the shared feature extraction layer can focus more on extracting features that are useful for abnormal risk classification, strengthen the feature extraction capability of the abnormal risk classification task, and improve its classification accuracy.

[0132] Step S2504: When the cosine similarity loss of the performance-degraded regression branch reaches a preset threshold, the parameter freezing state of all branches is released and the multi-task collaborative training mode is restored.

[0133] In this embodiment, a preset threshold is set, for example, when the cosine similarity loss of the performance degradation regression branch drops below 0.1. When the preset threshold is reached, it means that the performance degradation regression branch has achieved a good optimization effect. At this time, the parameter freezing state of all branches is released, including the previously frozen parameters of the real-time health status prediction branch and the abnormal risk classification branch. The multi-task collaborative training mode is restored so that each branch can update its parameters again according to its own error, and the shared feature extraction layer can perform comprehensive parameter adjustments based on the feedback information of all branches, thereby realizing collaborative optimization between multiple tasks and further improving the performance of the entire multi-task state evaluation model.

[0134] Step S260: Repeat the feature extraction and parameter updating process until the loss values of all branches converge to a preset threshold range, thereby generating a trained multi-task state evaluation model.

[0135] In this embodiment, the process of steps S230 to S250 can be continuously repeated, that is, each time the historical operation feature set is input into the shared feature extraction layer to generate a historical shared hidden feature vector, then it is input into each branch network to calculate the loss value, and then the parameter update amplitude of the shared feature extraction layer is adjusted according to the loss value. During this process, the loss values of each branch are continuously monitored. A threshold interval is preset. For example, for mean square error, cross entropy loss, and cosine similarity loss, the preset threshold interval is [0.05, 0.15]. When the mean square error of the real-time health status prediction branch, the cross entropy loss of the abnormal risk classification branch, and the cosine similarity loss of the performance degradation regression branch all converge within this preset threshold interval, it indicates that the model has learned sufficient feature information, and the prediction performance of each branch has reached a relatively stable and satisfactory level. At this time, the training process ends, and a trained multi-task state evaluation model is generated. This multi-task state evaluation model can be used to accurately evaluate the state of the operation feature set of the target servo drive device, generate real-time health status indicators, abnormal risk levels, and performance degradation trends, providing a reliable basis for subsequent equipment maintenance optimization strategies.

[0136] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of a servo drive system state monitoring system 100 based on multi-sensor fusion that can implement the idea of the present application provided in some embodiments of the present application. For example, the processor 120 can be used on the servo drive system state monitoring system 100 based on multi-sensor fusion and is used to execute the functions in the present application.

[0137] The servo drive system state monitoring system 100 based on multi-sensor fusion can be a general-purpose server or a special-purpose server, both of which can be used to implement the servo drive system state monitoring method based on multi-sensor fusion 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.

[0138] For example, a servo drive system state monitoring system 100 based on multi-sensor fusion 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 a disk, ROM, or RAM, or any combination thereof. Exemplarily, the servo drive system state monitoring system 100 based on multi-sensor fusion 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 servo drive system state monitoring system 100 based on multi-sensor fusion further includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0139] For ease of explanation, only one processor is described in the servo drive system state monitoring system 100 based on multi-sensor fusion. However, it should be noted that the servo drive system state monitoring system 100 based on multi-sensor fusion 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 performed or separately performed by multiple processors. For example, if the processor of the servo drive system state monitoring system 100 based on multi-sensor fusion performs step A and step B, it should be understood that step A and step B can also be jointly performed by two different processors or separately performed in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0140] 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 servo drive system state monitoring method based on multi-sensor fusion is implemented.

[0141] 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 monitoring the state of a servo drive system based on multi-sensor fusion, characterized in that, The method includes: Obtaining a multi-source sensor data set of a target servo drive device, where the multi-source sensor data set includes vibration waveform data, temperature gradient data, and current harmonic data; Performing time synchronization and amplitude calibration processing on the multi-source sensor data set to generate a standardized monitoring data set; Invoking a dynamic encoding model to extract operation characteristics of the standardized monitoring data set to generate an operation characteristic set of the target device. Among them, the dynamic encoding model performs dynamic correlation encoding on different types of characteristics through a multi-layer cross-attention mechanism. In each layer of the attention mechanism, the dynamic encoding model analyzes the correlation between two types of characteristics for weighted fusion and encoding; Based on a multi-task state evaluation model, performing state evaluation on the operation characteristic set to generate a real-time health status index, an abnormal risk level, and a performance degradation trend of the target device. Among them, the real-time health status index is determined by performing temporal dependence modeling on the shared hidden feature vector of the operation characteristic set based on a gated recurrent unit network. The health status label corresponding to the real-time health status index is labeled according to the ratio of the actual maintenance times to the standard maintenance period. The abnormal risk level is determined by focusing on key characteristics of the shared hidden feature vector of the operation characteristic set based on a multi-head self-attention mechanism. The abnormal risk level label corresponding to the abnormal risk level is graded according to the standard deviation of the feature fluctuation amplitude within N hours before the occurrence of the fault. The performance degradation trend is determined by performing non-linear mapping on the shared hidden feature vector of the operation characteristic set based on a radial basis function network. The performance degradation trend label corresponding to the performance degradation trend is linearly fitted according to the percentage of the output torque decay rate to the rated torque; Generating a device maintenance optimization strategy according to the real-time health status index, the abnormal risk level, and the performance degradation trend, and feeding back the device maintenance optimization strategy to an operation and maintenance control platform to trigger an active maintenance operation.

2. The method for monitoring the state of a servo drive system based on multi-sensor fusion according to claim 1, wherein, The performing time synchronization and amplitude calibration processing on the multi-source sensor data set to generate a standardized monitoring data set includes: Extracting a first sampling timestamp sequence of the vibration waveform data, a second sampling timestamp sequence of the temperature gradient data, and a third sampling timestamp sequence of the current harmonic data; Determining the maximum common time interval between the first sampling timestamp sequence, the second sampling timestamp sequence, and the third sampling timestamp sequence, and performing time axis alignment processing on the vibration waveform data, the temperature gradient data, and the current harmonic data based on a linear interpolation algorithm to generate a time synchronization data set; Invoking a preset set of sensor calibration coefficients to perform amplitude compensation processing on the vibration waveform data, the temperature gradient data, and the current harmonic data in the time synchronization data set respectively. Among them, the set of sensor calibration coefficients includes a frequency response compensation factor of a vibration sensor, a non-linear error correction factor of a temperature sensor, and a phase shift correction parameter of a current sensor; Normalize the vibration waveform data, temperature gradient data, and current harmonic data after amplitude compensation to generate a standardized monitoring data set.

3. The state monitoring method of the servo drive system based on multi-sensor fusion according to claim 1, characterized in that, Call the dynamic coding model to extract the operation characteristics of the standardized monitoring data set, generating an operation characteristics set of the target device, including: Extract time-domain characteristics from the vibration waveform data in the standardized monitoring data set to generate the vibration effective value, peak factor, and kurtosis coefficient; Perform wavelet packet decomposition on the vibration waveform data to extract the preset frequency band energy ratio characteristics and frequency band entropy value characteristics; Perform sliding window mean processing on the temperature gradient data to generate the temperature change rate characteristics and heat accumulation trend characteristics; Perform fast Fourier transform processing on the current harmonic data to extract the fundamental wave amplitude, harmonic distortion rate, and amplitude ratio of characteristic sub-harmonic components; Input the vibration effective value, peak factor, kurtosis coefficient, preset frequency band energy ratio characteristics, frequency band entropy value characteristics, temperature change rate characteristics, heat accumulation trend characteristics, fundamental wave amplitude, harmonic distortion rate, and amplitude ratio of characteristic sub-harmonic components into the dynamic coding model, and perform dynamic correlation coding through a multi-layer cross-attention mechanism to generate an operation characteristics set.

4. The state monitoring method of the servo drive system based on multi-sensor fusion according to claim 1, characterized in that Based on the multi-task state evaluation model, evaluate the state of the operation characteristics set to generate the real-time health status index, abnormal risk level, and performance degradation trend of the target device, including: Input the operation characteristics set into the shared feature extraction layer to generate a shared hidden feature vector; Input the shared hidden feature vector into the real-time health status prediction branch, abnormal risk classification branch, and performance degradation regression branch respectively; In the real-time health status prediction branch, perform temporal dependence modeling on the shared hidden feature vector based on the gated recurrent unit network and output the real-time health status index; In the abnormal risk classification branch, focus on the key features of the shared hidden feature vector based on the multi-head self-attention mechanism and output the probability distribution of the abnormal risk level; In the performance degradation regression branch, perform non-linear mapping on the shared hidden feature vector based on the radial basis function network and output the quantization value of the performance degradation trend; Dynamically adjust the weight allocation ratio of the shared feature extraction layer according to the output results of the real-time health status prediction branch, abnormal risk classification branch, and performance degradation regression branch to achieve multi-task collaborative optimization.

5. The state monitoring method of the servo drive system based on multi-sensor fusion according to claim 1, characterized in that Generate a device maintenance optimization strategy according to the real-time health status index, abnormal risk level, and performance degradation trend, including: Establish a threshold interval for the health status index, a threshold interval for the abnormal risk level, and a slope threshold for the performance degradation trend; When the real-time health status index is lower than the lower limit of the health status index threshold interval and the abnormal risk level exceeds the first risk threshold, generate an immediate shutdown and maintenance instruction; When the slope of the performance degradation trend exceeds the slope threshold and the abnormal risk level is within the second risk threshold interval, generate a preventive maintenance plan and spare part replacement suggestions; When the real-time health status index is in the middle of the health status index threshold interval and the abnormal risk level is lower than the third risk threshold, generate an operation parameter optimization and adjustment plan; Integrate the immediate shutdown for maintenance instruction, preventive maintenance plan, spare part replacement suggestion, and operation parameter optimization and adjustment plan into an equipment maintenance optimization strategy.

6. The method for monitoring the state of a servo drive system based on multi-sensor fusion according to claim 3, wherein The wavelet packet decomposition process for the vibration waveform data to extract the preset frequency band energy ratio feature and frequency band entropy value feature includes: Select the Daubechies wavelet basis function to perform 5-layer wavelet packet decomposition on the vibration waveform data to obtain 32 frequency band node coefficients; Calculate the energy value of each frequency band node, and divide the energy value of each frequency band by the total energy to obtain the preset frequency band energy ratio feature; Calculate the permutation entropy for the coefficient sequence of each frequency band node to obtain the frequency band entropy value feature; Screen the characteristic frequency bands with an energy ratio exceeding the set threshold, and add the corresponding frequency band entropy value feature and energy ratio feature to the operation feature set.

7. The state monitoring method of the servo drive system based on multi-sensor fusion according to claim 4, characterized in that The dynamic adjustment of the weight allocation ratio of the shared feature extraction layer according to the output results of the real-time health status prediction branch, abnormal risk classification branch, and performance degradation regression branch to achieve multi-task collaborative optimization includes: Calculate the prediction error rate of the real-time health status prediction branch, the cross-entropy loss value of the abnormal risk classification branch, and the mean square error of the performance degradation regression branch; Normalize the prediction error rate, cross-entropy loss value, and mean square error to obtain the weight adjustment factors for each branch; Dynamically update the channel attention weights of each convolutional kernel in the shared feature extraction layer according to the weight adjustment factors; When the weight adjustment factor of the abnormal risk classification branch exceeds the preset alarm threshold, freeze the parameter update of the performance degradation regression branch and prioritize the optimization of the abnormal risk classification task.

8. The state monitoring method of the servo drive system based on multi-sensor fusion according to claim 5, characterized in that, The generation of the preventive maintenance plan and spare part replacement suggestion includes: Calculate the predicted value of the remaining service life according to the slope of the performance degradation trend; Query the equipment maintenance knowledge graph to match the set of historical maintenance cases corresponding to the predicted value of the remaining service life; Extract the optimal maintenance time window, spare part model replacement record, and labor cost data from the historical maintenance cases; Based on the fuzzy decision-making algorithm, perform multi-objective optimization on the maintenance time window, spare part inventory status, and production plan to generate the time node of the preventive maintenance plan, the list of required spare parts, and the human resource allocation plan.

9. The state monitoring method of the servo drive system based on multi-sensor fusion according to claim 8, wherein The multi-objective optimization of the maintenance time window, spare part inventory status, and production plan based on the fuzzy decision-making algorithm to generate the time node of the preventive maintenance plan, the list of required spare parts, and the human resource allocation plan includes: Convert the maintenance time window into a time overlap conflict matrix, map the spare part inventory status to a spare part availability vector, and parse the production plan into a maintenance operation time constraint sequence; Construct a multi-objective optimization space based on the time overlap conflict matrix, spare part availability vector, and maintenance operation time constraint sequence, and use the Pareto front search algorithm to generate a set of candidate maintenance plans; Perform time conflict resolution processing on the set of candidate maintenance plans, eliminate the candidate plans that conflict with the maintenance operation time constraint sequence, and generate a set of conflict-free candidate maintenance plans; Sort the set of conflict-free candidate maintenance plans according to the spare part matching degree of the spare part availability vector, and select the top k candidate plans with the highest spare part matching degree; Input the top k candidate solutions into the fuzzy decision-making device, and generate the time nodes of the preventive maintenance plan, the list of required spare parts, and the human resource allocation plan according to the preset maintenance priority rules.

10. A servo drive system state monitoring system based on multi-sensor fusion, characterized in that, It 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 multi-sensor fusion-based servo drive system status monitoring method described in any one of claims 1-9 above.

Images