Soil remediation plant risk early warning method fusing multi-source data and carbon emission accounting

Abstract

The application provides a soil remediation plant risk early warning method fusing multi-source data and carbon emission accounting, the method comprises the following steps: obtaining plant multi-source time series data and carbon emission accounting data, and constructing a multi-dimensional state vector set at each sampling time; first layer quantization is performed on the state vector, the time change rate of each dimension is calculated and dimensionless, the first layer codebook is clustered to obtain the initial residual error; subsequent layers are weighted based on the data source contribution matrix, the codebook size is determined according to the distribution divergence, the codebook is generated layer by layer and the residual error is updated; after iteration to the maximum number of layers, the quantization path code is generated, the path probability is counted, the comprehensive risk score is calculated by combining the path probability reciprocal and the risk residual value, and the risk early warning is triggered when the threshold value is exceeded.

Description

Technical Field

[0001] This application belongs to the field of early warning, and in particular relates to a risk early warning method for soil remediation plants that integrates multi-source data and carbon emission accounting. Background Technology

[0002] With the continuous advancement of ecological environmental protection and green development concepts, soil remediation plants, in the process of treating contaminated soil, not only need to monitor various indicators such as equipment operating status, environmental parameters, and soil physicochemical properties in real time, but also need to routinely calculate energy consumption and carbon emissions during operation. Because soil remediation processes involve multiple stages such as crushing and screening, reagent dosing, heat treatment, exhaust gas treatment, and pollutant stabilization, the operating load of various equipment fluctuates significantly. Any abnormality in any stage can easily lead to risks such as secondary pollution, equipment failure, reduced remediation effectiveness, or excessive carbon emissions. Therefore, collecting and integrating multi-source status data during plant operation to achieve real-time monitoring and risk warning is a crucial means to ensure the safe, stable, and compliant operation of soil remediation plants.

[0003] When extracting features and reducing the dimensionality of high-dimensional state vectors, conventional methods struggle to adaptively adjust the representation granularity based on the actual distribution characteristics of the data. This can lead to the loss of subtle but risk-indicating anomalous fluctuations during the compression process, resulting in insufficient residual representation or inadequate anomaly detection sensitivity. Furthermore, existing early warning indicators fail to fully reflect the correlations between multi-source data and the differences in contributions from different data sources to the risk formation process. When facing highly concealed and complex potential risks, false alarms or missed alarms are likely to occur. Therefore, how to deeply explore and integrate the multidimensional features of multi-source time-series data and carbon emission accounting data, combine the differences in contributions from different data sources for refined quantitative representation, and construct a comprehensive risk assessment mechanism to achieve reliable risk early warning is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] To address the shortcomings of existing technologies, such as insufficient multi-source state data fusion representation, fixed feature quantification granularity, easy loss of abnormal fluctuation information, and low accuracy of risk warning, this invention provides, in a first aspect, a risk warning method for soil remediation plants that integrates multi-source data and carbon emission accounting, comprising the following steps: Acquire multi-source time-series data and carbon emission accounting data of the soil remediation plant over a period of time, and merge them into a set containing multi-dimensional state vectors at each time point; The set of multidimensional state vectors is quantized layer by layer. For the first layer of quantization, starting from the second sampling time, the time change rate of each dimension of each multidimensional state vector in the set is calculated and dimensionless, and combined into a joint feature vector. The joint feature vectors corresponding to each valid sampling time after the time change rate calculation are clustered to generate the first-level quantization codebook. For each joint feature vector, the nearest cluster center is selected as the first-level cluster center, and the codebook index is recorded. The difference is calculated as the initial quantization residual vector. For each subsequent quantization layer, the data source contribution matrix is ​​constructed sequentially. The target residual vector is obtained by weighting the quantization residual vector of the previous layer. The distribution divergence of all target residual vectors in the current layer is calculated to determine the size of the quantization codebook in the current layer. The target residual vectors are clustered to generate the quantization codebook. The nearest cluster center is selected, the codebook index is recorded, and the difference is calculated as the quantization residual vector of the next layer. The process iterates through subsequent quantization levels until the preset maximum number of levels is reached, using the quantization residual vector as the final residual vector. The codebook index is extracted and concatenated into a quantization path code, and the occurrence probability is calculated. The modulus is calculated to obtain the risk residual value. The comprehensive risk score is calculated by combining the risk residual value with the reciprocal of the occurrence probability. A risk warning is triggered when the score exceeds the alarm threshold.

[0005] Furthermore, the acquisition of multi-source time-series data and carbon emission accounting data of the soil remediation plant over a period of time is fused into a set containing multi-dimensional state vectors at each time point, including: The multi-source time-series data and the carbon emission accounting data are resampled at a fixed sampling frequency. Extract the original values ​​of each data item at each time point, and use the max-min normalization method based on the historical maximum and minimum values ​​of the corresponding parameters to map the original values ​​to the range of 0 to 1 to generate standardized data; For each time step, all standardized data at the same time step are sequentially concatenated and constructed into a one-dimensional row vector according to a fixed arrangement order. The one-dimensional row vectors corresponding to each time step constitute the set of the multi-dimensional state vectors.

[0006] Furthermore, the step of calculating the time change rate of each dimension of each multidimensional state vector in the set and performing dimensionless processing, starting from the second sampling time, includes: Starting from the second sampling time, obtain the values ​​of each dimension in the multidimensional state vector at the current time. Extract the values ​​of the same dimension from the multidimensional state vector corresponding to the previous sampling time. The difference is obtained by subtracting the value at the current time from the value at the previous sampling time. Divide the difference by the time interval constant between two adjacent sampling times to obtain the rate of change of time. The time change rate is mapped to the interval between 0 and 1 using a maximum and minimum value standardization method based on a preset interval limit of a historical sample database, thus completing the dimensionless processing.

[0007] Further, the first-level quantization codebook is generated by clustering the joint feature vectors corresponding to each valid sampling time for completing the calculation of the rate of change over time. For each joint feature vector, the nearest cluster center is selected as the first-level cluster center, and the codebook index is recorded, including: The K-means clustering algorithm is used to iteratively calculate the joint feature vectors corresponding to each effective sampling time for completing the calculation of the rate of change of time in order to calculate the cluster centers. Extract all cluster centers obtained after the clustering iteration process converges and use them as the first-level quantization codebook; Calculate the Euclidean distance between each joint feature vector and each cluster center in the first-level quantized codebook; Find the cluster center with the smallest Euclidean distance value as the first-level cluster center corresponding to the joint feature vector, and extract the index of the cluster center in the first-level quantized codebook as the codebook index.

[0008] Furthermore, the construction of the data source contribution matrix includes: The carbon emissions at the same effective sampling time are summed to obtain the comprehensive carbon emission accounting data sequence. The absolute value of the Pearson correlation coefficient between the data sequence corresponding to each quantitative residual dimension and the comprehensive carbon emission accounting data sequence within the same effective sampling time range is calculated, and the absolute value is used as the correlation degree. Extract the pre-configured constants corresponding to each quantized residual dimension as the preset importance; The data source contribution matrix is ​​a diagonal matrix. The elements on the main diagonal of the matrix are formed by sequentially arranging the products of the preset importance and correlation corresponding to each dimension of the quantized residual vector, and all elements on the non-main diagonal positions are zero.

[0009] Further, the step of calculating the distribution divergence of all target residual vectors at the current level, and determining the quantization codebook size at the current level accordingly, includes: For all target residual vectors in the current level, calculate the variance of each dimension; The variances of each dimension are weighted using a preset dimension normalization weighting coefficient, and the weighted variances of each dimension are summed. The sum is then determined as the distribution divergence. Multiply the distribution divergence by the preset codebook expansion constant to obtain the intermediate value of the codebook size; The intermediate value of the codebook size is rounded down to obtain the theoretical number of quantizations; The size of the current level of quantization codebook depends on the comparison between the theoretical number of quantizations and the preset minimum size constant: when the theoretical number of quantizations is greater than the preset minimum size constant, the theoretical number of quantizations is taken as the codebook size; Otherwise, the codebook size is taken as the preset minimum size constant.

[0010] Furthermore, the calculation of the comprehensive risk score by combining the risk residual value with the reciprocal of the probability of occurrence includes: The path rarity term is obtained by adding the occurrence probability to a preset minimum positive number and taking the reciprocal. The path rarity term is multiplied by a preset first weight coefficient to obtain the first scoring term; The risk residual value is multiplied by a preset second weighting coefficient to obtain a second scoring item; the first scoring item and the second scoring item are added together to obtain the comprehensive risk score. Wherein, both the preset first weight coefficient and the preset second weight coefficient are greater than zero.

[0011] On the other hand, the present invention also provides a soil remediation plant risk early warning system that integrates multi-source data and carbon emission accounting, including the following modules: The acquisition module is used to acquire multi-source time-series data and carbon emission accounting data of the soil remediation plant over a period of time, and integrate them into a set containing multi-dimensional state vectors at each time point; The computation module performs layer-by-layer quantization on the set of multidimensional state vectors. For the first layer of quantization, starting from the second sampling time, it calculates the time change rate of each dimension of each multidimensional state vector in the set and performs dimensionless processing, combining them into a joint feature vector. It then clusters the joint feature vectors corresponding to each valid sampling time after the time change rate calculation to generate the first-level quantization codebook. For each joint feature vector, it selects the nearest cluster center as the first-level cluster center and records the codebook index. Finally, it calculates the difference as the initial quantization residual vector. The generation module is used to sequentially construct the data source contribution matrix for each subsequent quantization layer, weight the quantization residual vector of the previous layer to obtain the target residual vector, calculate the distribution divergence of all target residual vectors in the current layer, and determine the quantization codebook size of the current layer accordingly; cluster the target residual vectors to generate the quantization codebook, select the nearest cluster center, record the codebook index, and calculate the difference as the quantization residual vector of the next layer. The early warning module is used to cyclically execute subsequent levels of quantization up to the preset maximum number of levels, using the quantization residual vector as the final residual vector; extract the codebook index and concatenate it into quantization path encoding, and calculate the occurrence probability; calculate the modulus to obtain the risk residual value; combine the risk residual value with the reciprocal of the occurrence probability to calculate the comprehensive risk score, and trigger a risk warning when it exceeds the alarm threshold.

[0012] Preferably, the acquisition of multi-source time-series data and carbon emission accounting data of the soil remediation plant over a period of time, and the fusion of these into a set containing multi-dimensional state vectors at each time point, includes: The multi-source time-series data and the carbon emission accounting data are resampled at a fixed sampling frequency. Extract the original values ​​of each data item at each time point, and use the max-min normalization method based on the historical maximum and minimum values ​​of the corresponding parameters to map the original values ​​to the range of 0 to 1 to generate standardized data; For each time step, all standardized data at the same time step are sequentially concatenated and constructed into a one-dimensional row vector according to a fixed arrangement order. The one-dimensional row vectors corresponding to each time step constitute the set of the multi-dimensional state vectors.

[0013] Preferably, the step of calculating the time change rate of each dimension of each multidimensional state vector in the set and performing dimensionless processing, starting from the second sampling time, includes: Starting from the second sampling time, obtain the values ​​of each dimension in the multidimensional state vector at the current time. Extract the values ​​of the same dimension from the multidimensional state vector corresponding to the previous sampling time. The difference is obtained by subtracting the value at the current time from the value at the previous sampling time. Divide the difference by the time interval constant between two adjacent sampling times to obtain the rate of change of time. The time change rate is mapped to the interval between 0 and 1 using a maximum and minimum value standardization method based on a preset interval limit of a historical sample database, thus completing the dimensionless processing.

[0014] Preferably, the step of clustering the joint feature vectors corresponding to each valid sampling time for completing the time change rate calculation to generate a first-level quantization codebook, and selecting the nearest cluster center as the first-level cluster center for each joint feature vector, and recording the codebook index, includes: The K-means clustering algorithm is used to iteratively calculate the joint feature vectors corresponding to each effective sampling time for completing the calculation of the rate of change of time in order to calculate the cluster centers. Extract all cluster centers obtained after the clustering iteration process converges and use them as the first-level quantization codebook; Calculate the Euclidean distance between each joint feature vector and each cluster center in the first-level quantized codebook; Find the cluster center with the smallest Euclidean distance value as the first-level cluster center corresponding to the joint feature vector, and extract the index of the cluster center in the first-level quantized codebook as the codebook index.

[0015] Preferably, constructing the data source contribution matrix includes: The carbon emissions at the same effective sampling time are summed to obtain the comprehensive carbon emission accounting data sequence. The absolute value of the Pearson correlation coefficient between the data sequence corresponding to each quantitative residual dimension and the comprehensive carbon emission accounting data sequence within the same effective sampling time range is calculated, and the absolute value is used as the correlation degree. Extract the pre-configured constants corresponding to each quantized residual dimension as the preset importance; The data source contribution matrix is ​​a diagonal matrix. The elements on the main diagonal of the matrix are formed by sequentially arranging the products of the preset importance and correlation corresponding to each dimension of the quantized residual vector, and all elements on the non-main diagonal positions are zero.

[0016] Preferably, the step of calculating the distribution divergence of all target residual vectors at the current level, and determining the quantization codebook size at the current level accordingly, includes: For all target residual vectors in the current level, calculate the variance of each dimension; The variances of each dimension are weighted using a preset dimension normalization weighting coefficient, and the weighted variances of each dimension are summed. The sum is then determined as the distribution divergence. Multiply the distribution divergence by the preset codebook expansion constant to obtain the intermediate value of the codebook size; The intermediate value of the codebook size is rounded down to obtain the theoretical number of quantizations; The size of the current level of quantization codebook depends on the comparison between the theoretical number of quantizations and the preset minimum size constant: when the theoretical number of quantizations is greater than the preset minimum size constant, the theoretical number of quantizations is taken as the codebook size; Otherwise, the codebook size is taken as the preset minimum size constant.

[0017] Preferably, the step of calculating the comprehensive risk score by combining the risk residual value with the reciprocal of the probability of occurrence includes: The path rarity term is obtained by adding the occurrence probability to a preset minimum positive number and taking the reciprocal. The path rarity term is multiplied by a preset first weight coefficient to obtain the first scoring term; The risk residual value is multiplied by a preset second weighting coefficient to obtain a second scoring item; the first scoring item and the second scoring item are added together to obtain the comprehensive risk score. Wherein, both the preset first weight coefficient and the preset second weight coefficient are greater than zero.

[0018] The present invention has the following advantages: By integrating multi-source time-series data with carbon emission accounting data, a multi-dimensional state vector is constructed to achieve a unified characterization of the operating status and carbon emission impact of soil remediation plants. In the quantization process, the first-level quantization extracts state change features based on the time rate of change and dimensionless results; subsequent levels use the data source contribution matrix to weight the residual vector and adaptively determine the codebook size based on the distribution divergence, thereby improving the ability to characterize subtle fluctuations in high-dimensional complex data and enhancing adaptability to changing working conditions. The multi-layer codebook index generates quantized path codes, and the comprehensive risk score is calculated by combining the reciprocal of the path occurrence probability with the final residual modulus. This transforms complex operational anomalies into comparable quantitative indicators, improves the sensitivity of identifying low-frequency anomalies, and enables reliable early warning of operational risks in soil remediation plants. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a risk early warning method for soil remediation plants that integrates multi-source data and carbon emission accounting; Figure 2 A flowchart illustrating the fusion process of multi-source data and carbon emission data; Figure 3 Flowchart for generating the contribution matrix. Detailed Implementation

[0020] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0021] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] In the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0023] References to "one embodiment" or "some embodiments" as described in this application specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in another embodiment," "in yet another embodiment," etc., appearing in different parts of this application specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0024] This invention proposes a risk early warning method for soil remediation plants that integrates multi-source data and carbon emission accounting. Figure 1 As shown, it includes the following steps: S1 acquires multi-source time-series data and carbon emission accounting data of the soil remediation plant over a period of time, and merges them into a set containing multi-dimensional state vectors at each time point.

[0025] Multi-source time-series data is collected through IoT sensors deployed at the soil remediation plant. This data includes soil temperature, soil pH, volatile organic compound (VOC) concentration, and target heavy metal ion concentration. The target heavy metal ion concentration is a pre-selected concentration; in this embodiment, lead ion concentration is used. Carbon emission accounting data is collected through the plant's energy management system. This data includes carbon emissions converted from equipment power consumption, carbon emissions from fuel consumption by excavation equipment, and carbon emissions converted from remediation agent consumption.

[0026] The data collection period is divided into discrete time steps of equal length. To address the inconsistency in sampling frequencies between multi-source time-series data and carbon emission accounting data, linear interpolation is used to align all data to the same discrete time step. At each discrete time step, the aligned soil temperature, soil pH, volatile organic compound concentration, target heavy metal ion concentration, and various carbon emissions are sequentially concatenated in a predetermined fixed order to form a multidimensional state vector for that moment. The multidimensional state vectors corresponding to all discrete time steps within the time period are collected to construct a set of multidimensional state vectors. The multi-source data and carbon emission data fusion process is as follows: Figure 2 As shown.

[0027] As an optional implementation, the acquisition of multi-source time-series data and carbon emission accounting data of the soil remediation plant over a period of time, and the fusion of these into a set containing multi-dimensional state vectors at each time point, includes: The multi-source time-series data and the carbon emission accounting data are resampled at a fixed sampling frequency. Extract the original values ​​of each data item at each time point, and use the max-min normalization method based on the historical maximum and minimum values ​​of the corresponding parameters to map the original values ​​to the range of 0 to 1 to generate standardized data; For each time step, all standardized data at the same time step are sequentially concatenated and constructed into a one-dimensional row vector according to a fixed arrangement order. The one-dimensional row vectors corresponding to each time step constitute the set of the multi-dimensional state vectors.

[0028] The data acquisition system is set to a fixed sampling interval of 10 minutes (1 / 6 of an hour) to synchronously extract multi-source time-series data and carbon emission accounting data from the sensor network and energy management system. The multi-source time-series data includes soil temperature, pH value, volatile organic compound concentration, and lead ion concentration. The carbon emission accounting data includes parameters such as carbon emissions converted from electricity consumption, carbon emissions converted from fuel consumption of excavating equipment, and carbon footprint of pesticide consumption. Assuming the raw value of the lead ion concentration acquired at a certain moment is 450 mg / kg, and the preset historical maximum value for this parameter is 1000 mg / kg, and the preset minimum value is 50 mg / kg, the standardized value of 0.421 is calculated using the maximum-minimum value normalization formula, mapping the parameter to the range of 0 to 1. For all other data, the historical maximum and minimum values ​​of the corresponding parameters are converted to standardized values ​​one by one using the same normalization method.

[0029] At each sampling time Following a pre-set fixed order of temperature, pH value, volatile organic compound concentration, lead ion concentration, carbon emissions from electricity, carbon emissions from fuel, and carbon emissions from chemicals, all standardized data calculated at that moment are sequentially arranged and pieced together. For example, [the following is a list of data points]. The standardized values ​​at each time point [0.552, 0.702, 0.218, 0.421, 0.613, 0.276, 0.128] are constructed as a one-dimensional row vector with 7 dimensions. As the sampling time window progresses, a total of 1008 row vectors corresponding to sampling times are acquired over a continuous week. These row vectors are stacked in chronological order to construct a two-dimensional data matrix of size 1008×7. Each row of this matrix is ​​an element in the set of multidimensional state vectors.

[0030] S2 performs layer-by-layer quantization on the set of multidimensional state vectors. For the first layer of quantization, starting from the second sampling time, the time change rate of each dimension of each multidimensional state vector in the set is calculated and dimensionless, and combined into a joint feature vector. The joint feature vectors corresponding to each valid sampling time after the time change rate calculation are clustered to generate the first-level quantization codebook. For each joint feature vector, the nearest cluster center is selected as the first-level cluster center, and the codebook index is recorded. The difference is calculated as the initial quantization residual vector.

[0031] The total number of quantization layers is set, and the first layer of quantization is executed. For the current multidimensional state vector in the multidimensional state vector set, the vector is subtracted from the previous multidimensional state vector by the corresponding dimension, and the result is divided by the time step to obtain the time change rate of each dimension at the current moment. The maximum-minimum normalization method is used to perform dimensionless processing on the time change rate of each dimension, mapping the processed values ​​to the interval between 0 and 1. The time change rate normalization uses the historical maximum and minimum values ​​of the time change rate of each dimension, which are fixed and stored in a preset historical sample library, as the normalization boundary to ensure that the real-time data is consistent with the quantization benchmark in the historical sample library. These dimensionless time change rates are arranged and combined sequentially according to the dimensional order of the multidimensional state vector to form a joint feature vector.

[0032] The joint feature vectors corresponding to all sampling moments for which the rate of change of time has been calculated are input into a K-means clustering algorithm model based on Euclidean distance. The number of clusters in the first level is preset to a fixed positive integer. Through iterative optimization, the sum of squares within each cluster is minimized. After clustering, a first-level quantized codebook is obtained, which consists of the center vectors of all clusters. This first-level quantized codebook is trained offline during the historical sample database establishment phase and is permanently stored along with the codebook version number, cluster center number, and normalized boundary parameters. During the real-time risk warning phase, the same version of the first-level quantized codebook is used for nearest-center matching.

[0033] For each joint feature vector, calculate the Euclidean distance between the vector and all center vectors in the first-level quantization codebook. Select the center vector with the smallest Euclidean distance as the first-level cluster center of the joint feature vector, and record the position number of this cluster center in the codebook as the first-level codebook index. Subtract the first-level cluster center vector corresponding to the vector from the joint feature vector, and the resulting difference vector is used as the initial quantization residual vector at the current time.

[0034] As an optional implementation, the step of calculating the time change rate of each dimension of each multidimensional state vector in the set and performing dimensionless processing, starting from the second sampling time, includes: Starting from the second sampling time, obtain the values ​​of each dimension in the multidimensional state vector at the current time. Extract the values ​​of the same dimension from the multidimensional state vector corresponding to the previous sampling time. The difference is obtained by subtracting the value at the current time from the value at the previous sampling time. Divide the difference by the time interval constant between two adjacent sampling times to obtain the rate of change of time. The time change rate is mapped to the interval between 0 and 1 using a maximum and minimum value standardization method based on a preset interval limit of a historical sample database, thus completing the dimensionless processing.

[0035] For the constructed multidimensional state vector set, data from adjacent time points are extracted sequentially according to the time series for change feature analysis. Assume the current time... The first dimension of the multidimensional state vector, for example, the standardized value of temperature, is 0.582, while the value of the previous sampling time is... The value for the same dimension is 0.552. The calculated difference between the two is 0.030. Given that the system is set to set the time interval constant Δt between two adjacent sampling times to be 10 minutes, dividing the difference 0.030 by this time interval constant 10 yields the time change rate of the temperature dimension v = 0.003 / minute.

[0036] After calculating the rate of change for all dimensions, the historical sample database of the time-varying rate for that dimension is read and normalized according to the preset interval boundaries. Assuming the historical minimum value of the time-varying rate for the first dimension is -0.010 / minute and the maximum value is 0.015 / minute, the dimensionless result is calculated as 0.520 using the maximum-minimum-maximum normalization formula. The above operations of taking values, calculating differences, dividing by the time interval, and normalizing are repeated sequentially for each dimension in the vector, representing the state changes between adjacent sampling times as dimensionless vector features of process fluctuations. These dimensionless rates of change are concatenated in the original dimensional order to generate a continuous feature array representing the evolution rate of factory operation.

[0037] As an optional implementation, the step of clustering the joint feature vectors corresponding to each valid sampling time for completing the time change rate calculation to generate a first-level quantization codebook, and selecting the nearest cluster center as the first-level cluster center for each joint feature vector, and recording the codebook index, includes: The K-means clustering algorithm is used to iteratively calculate the joint feature vectors corresponding to each effective sampling time for completing the calculation of the rate of change of time in order to calculate the cluster centers. Extract all cluster centers obtained after the clustering iteration process converges and use them as the first-level quantization codebook; Calculate the Euclidean distance between each joint feature vector and each cluster center in the first-level quantized codebook; Find the cluster center with the smallest Euclidean distance value as the first-level cluster center corresponding to the joint feature vector, and extract the index of the cluster center in the first-level quantized codebook as the codebook index.

[0038] The K-means K-Means clustering algorithm was used to input a total of 1007 effective joint feature vectors formed by sampling over a week into the model. Each joint feature vector is 7-dimensional and contains the dimensionless time change rate corresponding to the 7 data dimensions.

[0039] The number of clusters for the K-means algorithm is set to K=64. Initial centroids are randomly selected using the K-Means++ strategy. After training is completed during the historical sample database establishment phase, the first-level codebook version is permanently saved. Through multiple rounds of iteration, the distance from each vector to each centroid is calculated, and the cluster centroid positions are updated again until the moving mean square error of the centroid positions is less than the preset convergence tolerance of 1×10⁻⁶. -4 The maximum number of iterations may be reached, up to 300. After convergence, the 64 7-dimensional centroid vectors of the output are extracted and saved as a first-level quantization codebook matrix of size 64×7.

[0040] For any given joint feature vector The vector and the 1st to 64th cluster centers in the codebook matrix are calculated sequentially by vector subtraction and summation of squares. The Euclidean distance between them. If, after traversing and calculating, it is found that this vector is close to the 15th cluster center... Euclidean distance between =0.087 is the smallest among all 64 distance values, thus determining... This is the first-level cluster center for this feature vector. The integer index "15" is recorded as the first-level codebook index value corresponding to the feature at this moment.

[0041] Using the original joint feature vector Subtract the cluster center corresponding to the vector This yields a residual vector that is also 7-dimensional.

[0042] S3: For each subsequent quantization layer, construct the data source contribution matrix in sequence, and weight the quantization residual vector of the previous layer to obtain the target residual vector; calculate the distribution divergence of all target residual vectors in the current layer, and determine the quantization codebook size of the current layer accordingly; cluster the target residual vectors to generate the quantization codebook, select the nearest cluster center, record the codebook index, and calculate the difference as the quantization residual vector of the next layer.

[0043] The process then proceeds to the second and deeper levels of quantization, extracting the quantization residual vectors corresponding to each effective sampling time from the previous level.

[0044] When constructing the data source contribution matrix, the carbon emissions corresponding to the power consumption of equipment, the carbon emissions corresponding to the fuel consumption of excavating equipment, and the carbon emissions corresponding to the consumption of remediation agents at the same valid sampling time are summed to obtain a comprehensive carbon emission accounting data sequence. Since the quantization residual vector is generated starting from the second sampling time, the comprehensive carbon emission accounting data sequence synchronously selects data within the same valid sampling time range, ensuring that the quantization residual sequence and the comprehensive carbon emission accounting data sequence are consistent in time position and sequence length.

[0045] The absolute value of the Pearson correlation coefficient between the data sequences corresponding to each dimension of the previous-level quantized residual vector and the comprehensive carbon emission accounting data sequence is calculated. This absolute value is then multiplied by the preset importance corresponding to each dimension to obtain the data source contribution weight for each dimension. These weights are used as diagonal elements to construct a data source contribution matrix. The quantized residual vector at the corresponding time point of the previous level is multiplied by this data source contribution matrix to weight the residuals of different dimensions, thus obtaining the target residual vector for the current level.

[0046] For all target residual vectors at the current level, calculate the variance of each dimension, and then use preset dimension normalization weighting coefficients to weight and accumulate the variances of each dimension. The weighted sum is used as the distribution divergence of the target residual vector at the current level. The contribution matrix generation process is as follows: Figure 3 As shown.

[0047] Multiply the distribution divergence by the preset codebook expansion constant to obtain the intermediate value of the codebook size; perform floor processing on the intermediate value of the codebook size to obtain the theoretical number of quantizations; compare the theoretical number of quantizations with the preset minimum size constant, and determine the larger value of the two as the codebook size of the current level of quantization.

[0048] During the historical sample database establishment phase, K-means clustering is performed on all target residual vectors at the current level using a defined number of clusters. After iterative convergence, the quantization codebook for the current level is generated, and the codebook version number, codebook size, cluster center vector, and index number are saved according to the level. During the real-time alert phase, a fixed version codebook consistent with the historical sample database is invoked. For each target residual vector, its Euclidean distance to each cluster center in the current level quantization codebook is calculated. The cluster center with the smallest Euclidean distance is selected, and its position number in the current level codebook is recorded as the current level codebook index. The target residual vector is subtracted from the selected cluster center vector, and the resulting difference vector is used as the quantization residual vector passed to the next level.

[0049] As an optional implementation, constructing the data source contribution matrix includes: The carbon emissions at the same effective sampling time are summed to obtain the comprehensive carbon emission accounting data sequence. The absolute value of the Pearson correlation coefficient between the data sequence corresponding to each quantitative residual dimension and the comprehensive carbon emission accounting data sequence within the same effective sampling time range is calculated, and the absolute value is used as the correlation degree. Extract the pre-configured constants corresponding to each quantized residual dimension as the preset importance; The data source contribution matrix is ​​a diagonal matrix. The elements on the main diagonal of the matrix are formed by sequentially arranging the products of the preset importance and correlation corresponding to each dimension of the quantized residual vector, and all elements on the non-main diagonal positions are zero.

[0050] For the 7-dimensional quantized residual vector sequence generated after the first-level quantization, historical data sequences for each independent dimension are extracted within a global time window, along with comprehensive carbon emission accounting data sequences within the same valid time range. The covariance between the i-th quantized residual dimension sequence and the comprehensive carbon emission accounting data sequence is calculated and divided by the product of their standard deviations to obtain the Pearson correlation coefficient. Assuming the calculated correlation coefficient for the temperature residual dimension is -0.85, the absolute value of this correlation coefficient is used to obtain the correlation degree of that dimension. =0.85; the correlation coefficient for the pH residual dimension is 0.42, thus obtaining the correlation degree. =0.42.

[0051] The dimensionless constants for importance of each dimension are pre-configured based on historical operating data, process expert experience, and the risk impact level corresponding to each monitoring indicator, and stored in the system parameter table. When constructing the data source contribution matrix, the corresponding dimensionless constants for importance are read according to the fixed dimension order of the multidimensional state vector. For example, the importance of the temperature dimension is preset. =1.2, pH value dimension =0.8. Multiplying the two by a scalar yields the weighted values ​​for each dimension: the product for temperature is 1.02, and the product for pH is 0.336. After calculating the products for all m=7 dimensions, a 7×7 zero matrix is ​​constructed. The calculated seven product values ​​are then filled into the positions corresponding to the main diagonal of this matrix in dimensional order using the diagonalization function `diag`. Elements in the off-diagonal positions are set to zero, thus constructing a diagonal matrix W.

[0052] As an optional implementation, the step of calculating the distribution divergence of all target residual vectors at the current level, and determining the quantization codebook size at the current level accordingly, includes: For all target residual vectors in the current level, calculate the variance of each dimension; The variances of each dimension are weighted using a preset dimension normalization weighting coefficient, and the weighted variances of each dimension are summed. The sum is then determined as the distribution divergence. Multiply the distribution divergence by the preset codebook expansion constant to obtain the intermediate value of the codebook size; The intermediate value of the codebook size is rounded down to obtain the theoretical number of quantizations; The size of the current level of quantization codebook depends on the comparison between the theoretical number of quantizations and the preset minimum size constant: when the theoretical number of quantizations is greater than the preset minimum size constant, the theoretical number of quantizations is taken as the codebook size; Otherwise, the codebook size is taken as the preset minimum size constant.

[0053] When calculating the size of the second-level quantized codebook, a matrix is ​​constructed by extracting all 1007 valid target weighted residual vectors generated in this layer. For this matrix's seven data dimensions (seven column vectors), the discrete variance of each dimension is calculated using the standard deviation formula. The dimension normalization weight coefficients corresponding to the variances of each dimension are read, and the variance values ​​of all seven dimensions are weighted and summed. Assuming the sum of all dimension variances is 0.358, this value is set as the dispersion D = 0.358 of the current level residual vector set.

[0054] After obtaining the distribution divergence, a fixed codebook expansion constant C=120.0 and a preset minimum size constant are obtained. =16. According to the formula =⌊0.358×120.0⌋, thus obtaining the theoretical quantization number. =42. Since 42>16, this satisfies... The size of the current level quantization codebook is determined to be N=42. If, at deeper levels, the residual decreases rapidly, causing the divergence to decrease, the obtained... If the value is 12, the number of quantizations is determined to be less than or equal to 12. The threshold is set to N=16 to ensure that the extremely small residual space still has the most basic representation granularity, and this value is used as the number of clusters to start K-means clustering at this level.

[0055] S4, cyclically execute subsequent levels of quantization up to the preset maximum number of levels, and use the quantization residual vector as the final residual vector; extract the codebook index and concatenate it into quantization path encoding, and count the occurrence probability; calculate the modulus to obtain the risk residual value; combine the risk residual value with the reciprocal of the occurrence probability to calculate the comprehensive risk score, and trigger a risk warning when it exceeds the alarm threshold.

[0056] Using the calculated residual vector of the next level of quantization as input, the process of constructing the data source contribution matrix and subsequent clustering and quantization operations is repeated until the current quantization level equals the initially set total number of quantization levels. At this point, the difference vector obtained from the last calculation is used as the final residual vector for that moment.

[0057] Following the quantization level order, the codebook indices of each level are arranged sequentially and converted into string format. These indices are then concatenated using level identifiers and delimiters to generate unambiguous quantization path codes. For example, a four-level quantization path can be represented as "L1:015|L2:007|L3:003|L4:011". The quantization path codes in a preset historical sample library consistent with the current codebook version are traversed. The frequency of occurrence of the current quantization path code in the preset historical sample library is counted, and this frequency is divided by the total number of samples in the preset historical sample library to obtain the probability of occurrence of the quantization path code.

[0058] The sum of squares of all elements in the final residual vector is calculated, and the square root of the sum is taken to obtain the Euclidean modulus. This modulus is used as the risk residual value. The risk residual value is used to characterize the residual strength of the current state that is not fully matched by the codebook after multi-layer quantization representation.

[0059] Set residual weight coefficients and probability weight coefficients, add the occurrence probability to a preset minimum positive number and take the reciprocal to obtain the path rarity item; multiply the path rarity item by a preset first weight coefficient to obtain a first scoring item, multiply the risk residual value by a preset second weight coefficient to obtain a second scoring item, and add the first scoring item and the second scoring item to obtain a comprehensive risk score.

[0060] An alarm threshold is pre-set in the factory's control system, and the overall risk score is compared with the alarm threshold in real time. When the overall risk score exceeds the alarm threshold, the control system generates an alarm command, activates the audible and visual alarms in the factory to flash and sound, and displays an alarm pop-up on the monitoring center screen containing the overall risk score, the corresponding sampling time, and the alarm status, thus completing the risk warning.

[0061] As an optional implementation, the step of calculating the comprehensive risk score by combining the risk residual value with the reciprocal of the probability of occurrence includes: The path rarity term is obtained by adding the occurrence probability to a preset minimum positive number and taking the reciprocal. The path rarity term is multiplied by a preset first weight coefficient to obtain the first scoring term; The risk residual value is multiplied by a preset second weighting coefficient to obtain a second scoring item; the first scoring item and the second scoring item are added together to obtain the comprehensive risk score. Wherein, both the preset first weight coefficient and the preset second weight coefficient are greater than zero.

[0062] When progressive quantization reaches the preset final layer, such as the fourth layer, for the current multidimensional state vector, the codebook indices obtained from each layer are sequentially concatenated to generate a quantization path code. This quantization path code is generated as a string with a level separator or in a fixed-length, zero-padded form, and is stored and retrieved together with the codebook version number. The ratio of the number of occurrences of the specific code matching the preset historical sample library to the total number of historical samples is calculated. Assume that the probability of a quantization path corresponding to an abnormally high energy consumption condition is P=0.002 at the current moment. A preset minimum constant to prevent division by zero is obtained. =1×10 -5 P and The sum is 0.00201, and the reciprocal of the result is approximately 497.51. This value amplifies the contribution weight of rare working conditions to the overall score.

[0063] Calculate the magnitude of the final residual vector formed after the current multidimensional state undergoes quantization at all levels. Assume the risk residual value calculated using the vector norm formula is R = 0.150. Obtain the preset first weighting coefficient α = 0.1 and the preset second weighting coefficient β = 100.0. Calculate the comprehensive risk score according to the formula: the score component corresponding to the rarity of the state anomaly is calculated to be 49.751; then calculate the component representing the intensity of the abnormal residuals not interpreted by the codebook in the weighted quantization space as 15.0. Add the two scalars to obtain the comprehensive risk score S = 64.751 at the current moment. When this score exceeds the set safety alarm threshold, such as 60.0, trigger an alarm action and output a risk warning log.

[0064] This invention also proposes a risk early warning system for soil remediation plants that integrates multi-source data and carbon emission accounting, comprising the following modules: The acquisition module is used to acquire multi-source time-series data and carbon emission accounting data of the soil remediation plant over a period of time, and integrate them into a set containing multi-dimensional state vectors at each time point; The computation module performs layer-by-layer quantization on the set of multidimensional state vectors. For the first layer of quantization, starting from the second sampling time, it calculates the time change rate of each dimension of each multidimensional state vector in the set and performs dimensionless processing, combining them into a joint feature vector. It then clusters the joint feature vectors corresponding to each valid sampling time after the time change rate calculation to generate the first-level quantization codebook. For each joint feature vector, it selects the nearest cluster center as the first-level cluster center and records the codebook index. Finally, it calculates the difference as the initial quantization residual vector. The generation module is used to sequentially construct the data source contribution matrix for each subsequent quantization layer, weight the quantization residual vector of the previous layer to obtain the target residual vector, calculate the distribution divergence of all target residual vectors in the current layer, and determine the quantization codebook size of the current layer accordingly; cluster the target residual vectors to generate the quantization codebook, select the nearest cluster center, record the codebook index, and calculate the difference as the quantization residual vector of the next layer. The early warning module is used to cyclically execute subsequent levels of quantization up to the preset maximum number of levels, using the quantization residual vector as the final residual vector; extract the codebook index and concatenate it into quantization path encoding, and calculate the occurrence probability; calculate the modulus to obtain the risk residual value; combine the risk residual value with the reciprocal of the occurrence probability to calculate the comprehensive risk score, and trigger a risk warning when it exceeds the alarm threshold.

[0065] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for risk early warning of soil remediation plants that integrates multi-source data and carbon emission accounting, characterized in that, Includes the following steps: Acquire multi-source time-series data and carbon emission accounting data of the soil remediation plant over a period of time, and merge them into a set containing multi-dimensional state vectors at each time point; The set of multidimensional state vectors is quantized layer by layer. For the first layer of quantization, starting from the second sampling time, the time change rate of each dimension of each multidimensional state vector in the set is calculated and dimensionless, and combined into a joint feature vector. The joint feature vectors corresponding to each valid sampling time after the time change rate calculation are clustered to generate the first-level quantization codebook. For each joint feature vector, the nearest cluster center is selected as the first-level cluster center, and the codebook index is recorded. The difference is calculated as the initial quantization residual vector. For each subsequent quantization layer, the data source contribution matrix is ​​constructed sequentially, and the target residual vector is obtained by weighting the quantization residual vector of the previous layer. Calculate the distribution divergence of all target residual vectors at the current level, and determine the size of the quantization codebook at the current level accordingly; The cluster target residual vector is used to generate a quantization codebook. The nearest cluster center is selected, the codebook index is recorded, and the difference is calculated as the quantization residual vector for the next level. The process is repeated until the preset maximum number of levels of quantization is reached, and the quantized residual vector is used as the final residual vector. Extract the codebook index and concatenate it into a quantized path code, then calculate the probability of its occurrence. The risk residual value is obtained by calculating the modulus. The comprehensive risk score is calculated by combining the risk residual value with the reciprocal of the probability of occurrence. A risk warning is triggered when the score exceeds the alarm threshold.

2. The method according to claim 1, characterized in that, The process involves acquiring multi-source time-series data and carbon emission accounting data from the soil remediation plant over a period of time, and fusing them into a set containing multi-dimensional state vectors for each time point, including: The multi-source time-series data and the carbon emission accounting data are resampled at a fixed sampling frequency. Extract the original values ​​of each data item at each time point, and use the max-min normalization method based on the historical maximum and minimum values ​​of the corresponding parameters to map the original values ​​to the range of 0 to 1 to generate standardized data; For each time step, all standardized data at the same time step are sequentially concatenated and constructed into a one-dimensional row vector according to a fixed arrangement order. The one-dimensional row vectors corresponding to each time step constitute the set of the multi-dimensional state vectors.

3. The method according to claim 1, characterized in that, Starting from the second sampling time, the process of calculating the time change rate of each dimension of each multidimensional state vector in the set and performing dimensionless processing includes: Starting from the second sampling time, obtain the values ​​of each dimension in the multidimensional state vector at the current time. Extract the values ​​of the same dimension from the multidimensional state vector corresponding to the previous sampling time. The difference is obtained by subtracting the value at the current time from the value at the previous sampling time. Divide the difference by the time interval constant between two adjacent sampling times to obtain the rate of change of time. The time change rate is mapped to the interval between 0 and 1 using a maximum and minimum value standardization method based on a preset interval limit of a historical sample database, thus completing the dimensionless processing.

4. The method according to claim 3, characterized in that, The first-level quantization codebook is generated by clustering the joint feature vectors corresponding to each valid sampling time for completing the calculation of the rate of change of time. For each joint feature vector, the nearest cluster center is selected as the first-level cluster center, and the codebook index is recorded, including: The K-means clustering algorithm is used to iteratively calculate the joint feature vectors corresponding to each effective sampling time for completing the calculation of the rate of change of time in order to calculate the cluster centers. Extract all cluster centers obtained after the clustering iteration process converges and use them as the first-level quantization codebook; Calculate the Euclidean distance between each joint feature vector and each cluster center in the first-level quantized codebook; Find the cluster center with the smallest Euclidean distance value as the first-level cluster center corresponding to the joint feature vector, and extract the index of the cluster center in the first-level quantized codebook as the codebook index.

5. The method according to claim 1, characterized in that, The construction of the data source contribution matrix includes: The carbon emissions at the same effective sampling time are summed to obtain the comprehensive carbon emission accounting data sequence. The absolute value of the Pearson correlation coefficient between the data sequence corresponding to each quantitative residual dimension and the comprehensive carbon emission accounting data sequence within the same effective sampling time range is calculated, and the absolute value is used as the correlation degree. Extract the pre-configured constants corresponding to each quantized residual dimension as the preset importance; The data source contribution matrix is ​​a diagonal matrix. The elements on the main diagonal of the matrix are formed by sequentially arranging the products of the preset importance and correlation corresponding to each dimension of the quantized residual vector, and all elements on the non-main diagonal positions are zero.

6. The method according to claim 1, characterized in that, The calculation of the distribution divergence of all target residual vectors at the current level, and the determination of the quantization codebook size at the current level based on this divergence, includes: For all target residual vectors in the current level, calculate the variance of each dimension; The variances of each dimension are weighted using a preset dimension normalization weighting coefficient, and the weighted variances of each dimension are summed. The sum is then determined as the distribution divergence. Multiply the distribution divergence by the preset codebook expansion constant to obtain the intermediate value of the codebook size; The intermediate value of the codebook size is rounded down to obtain the theoretical number of quantizations; The size of the current level of quantization codebook depends on the comparison between the theoretical number of quantizations and the preset minimum size constant: when the theoretical number of quantizations is greater than the preset minimum size constant, the theoretical number of quantizations is taken as the codebook size; Otherwise, the codebook size is taken as the preset minimum size constant.

7. The method according to claim 1, characterized in that, The calculation of the comprehensive risk score by combining the risk residual value with the reciprocal of the probability of occurrence includes: The path rarity term is obtained by adding the occurrence probability to a preset minimum positive number and taking the reciprocal. The path rarity term is multiplied by a preset first weight coefficient to obtain the first scoring term; The risk residual value is multiplied by a preset second weighting coefficient to obtain a second scoring item; the first scoring item and the second scoring item are added together to obtain the comprehensive risk score. Wherein, both the preset first weight coefficient and the preset second weight coefficient are greater than zero.

8. A risk early warning system for soil remediation plants that integrates multi-source data and carbon emission accounting, characterized in that, Includes the following modules: The acquisition module is used to acquire multi-source time-series data and carbon emission accounting data of the soil remediation plant over a period of time, and integrate them into a set containing multi-dimensional state vectors at each time point; The computation module performs layer-by-layer quantization on the set of multidimensional state vectors. For the first layer of quantization, starting from the second sampling time, it calculates the time change rate of each dimension of each multidimensional state vector in the set and performs dimensionless processing, combining them into a joint feature vector. It then clusters the joint feature vectors corresponding to each valid sampling time after the time change rate calculation to generate the first-level quantization codebook. For each joint feature vector, it selects the nearest cluster center as the first-level cluster center and records the codebook index. Finally, it calculates the difference as the initial quantization residual vector. The generation module is used to sequentially construct the data source contribution matrix for each subsequent quantization layer and to weight the quantization residual vector of the previous layer to obtain the target residual vector. Calculate the distribution divergence of all target residual vectors at the current level, and determine the size of the quantization codebook at the current level accordingly; The cluster target residual vector is used to generate a quantization codebook. The nearest cluster center is selected, the codebook index is recorded, and the difference is calculated as the quantization residual vector for the next level. The early warning module is used to repeatedly execute subsequent levels of quantization up to the preset maximum number of levels, and use the quantized residual vector as the final residual vector; Extract the codebook index and concatenate it into a quantized path code, then calculate the probability of its occurrence. The risk residual value is obtained by calculating the modulus. The comprehensive risk score is calculated by combining the risk residual value with the reciprocal of the probability of occurrence. A risk warning is triggered when the score exceeds the alarm threshold.

9. The system according to claim 8, characterized in that, The process involves acquiring multi-source time-series data and carbon emission accounting data from the soil remediation plant over a period of time, and fusing them into a set containing multi-dimensional state vectors for each time point, including: The multi-source time-series data and the carbon emission accounting data are resampled at a fixed sampling frequency. Extract the original values ​​of each data item at each time point, and use the max-min normalization method based on the historical maximum and minimum values ​​of the corresponding parameters to map the original values ​​to the range of 0 to 1 to generate standardized data; For each time step, all standardized data at the same time step are sequentially concatenated and constructed into a one-dimensional row vector according to a fixed arrangement order. The one-dimensional row vectors corresponding to each time step constitute the set of the multi-dimensional state vectors.

10. The system according to claim 8, characterized in that, Starting from the second sampling time, the process of calculating the time change rate of each dimension of each multidimensional state vector in the set and performing dimensionless processing includes: Starting from the second sampling time, obtain the values ​​of each dimension in the multidimensional state vector at the current time. Extract the values ​​of the same dimension from the multidimensional state vector corresponding to the previous sampling time. The difference is obtained by subtracting the value at the current time from the value at the previous sampling time. Divide the difference by the time interval constant between two adjacent sampling times to obtain the rate of change of time. The time change rate is mapped to the interval between 0 and 1 using a maximum and minimum value standardization method based on a preset interval limit of a historical sample database, thus completing the dimensionless processing.

Images